AI Development English

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 01: Speaking the AI Development Stack
Original fictional case: A document assistant displays the wrong refund deadline. Retrieval returned the current policy, which states 30 days, but the generated answer says 60. The interface displays the generated text unchanged. Priya owns the model integration; Alex manages the product. A demonstration is scheduled for Friday. No cause has been confirmed, and Alex cannot approve a production release alone.
Language workshop: The answer is wrong, but which layer failed?
Communication goal: Distinguish components and ask a diagnostic question without assigning blame.
Language guidance: A trace records events; a diagnosis explains them. A wrong answer identifies a symptom, not automatically a model defect. Keep retrieval output, model output, and displayed text separate.
Useful expressions (preserve the supplied facts): Which component first shows the wrong value? / The retrieved passage says 30 days; the generated answer says 60. / That locates the mismatch, but it does not establish the cause. / Can you attach the request ID and the redacted trace? / Product approval and technical diagnosis are different decisions. / Can we reproduce it with the same payload and configuration? / By "the AI," do you mean the model or the whole application? / I will send the evidence and name the unresolved question. / Does the interface display the response without changing it? / Let us inspect the boundary before assigning ownership. / Change one component while keeping the others fixed. / We have an observation, not a verified root cause. / Does this test depend on the policy index being current? / Please retain the configuration used for that request. / I can commit to an update, not to a fix by that time. / The demonstration remains scheduled; release approval is separate.
Vocabulary: application stack: The components that together deliver an application. / retrieval: Finding source material relevant to a request. / language model: A model that processes and generates language sequences. / user interface: The part of a system a user sees and operates. / orchestration: Coordination of model, retrieval, and tool steps. / inference: Running a trained model to produce an output. / endpoint: An address through which a service accepts requests. / payload: The data carried in a request or response. / trace: A linked record of events in a request's path. / dependency: A component or condition another part relies on. / upstream: Earlier in the relevant processing path. / downstream: Later in the relevant processing path. / symptom: An observed sign of a problem, not its cause. / root cause: The underlying cause established by investigation. / reproduction: Recreating a reported result under stated conditions. / request ID: An identifier used to locate a particular request. / schema: A defined structure for data and its fields. / tool call: A model-requested invocation of an available tool. / API: Application programming interface; a defined software interaction boundary. / contract: Agreed expectations between software components. / integration: Connecting components into a working system. / observability: The ability to infer system behavior from available signals. / triage: Initial assessment that determines priority and ownership. / handoff: Transfer of work and its relevant context.
Speaking task: Check the dialogue, then read the 30-day source and 60-day generated answer as separate observations. Locate the mismatch before assigning its cause. Switch roles. Ask for D-184, the redacted trace and the request configuration; retain the Thursday findings update without promising a fix. Complete the citation-link exchange. State that the generated response contains the citation while the interface omits it, then hand over the exact request.
Partner: Priya: AI integration engineer / Alex: Product manager
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A support assistant generates a correct answer with a citation ID. The interface omits the citation link. The display team needs the exact recorded request, not an assumption that retrieval failed.
Model for comparison AFTER my attempt, not a response to give me first: Alex: All right. I will describe the customer-visible problem and keep the demonstration decision separate. What should your update include for that decision?
Priya: The reproduction result, affected cases, remaining uncertainty, and a proposed next step. I will preserve the configuration so the comparison is meaningful.
Alex: That works. Please send it by three on Thursday, even if the investigation is incomplete. I will confirm who can decide about Friday.
Priya: Agreed. The handoff will say what we know, what we do not know, and which decision still needs an owner.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 01: Speaking the AI Development Stack
Original fictional case: A document assistant displays the wrong refund deadline. Retrieval returned the current policy, which states 30 days, but the generated answer says 60. The interface displays the generated text unchanged. Priya owns the model integration; Alex manages the product. A demonstration is scheduled for Friday. No cause has been confirmed, and Alex cannot approve a production release alone.
Language workshop: The answer is wrong, but which layer failed?
Communication goal: Distinguish components and ask a diagnostic question without assigning blame.
Language guidance: A trace records events; a diagnosis explains them. A wrong answer identifies a symptom, not automatically a model defect. Keep retrieval output, model output, and displayed text separate.
Useful expressions (preserve the supplied facts): Which component first shows the wrong value? / The retrieved passage says 30 days; the generated answer says 60. / That locates the mismatch, but it does not establish the cause. / Can you attach the request ID and the redacted trace? / Product approval and technical diagnosis are different decisions. / Can we reproduce it with the same payload and configuration? / By "the AI," do you mean the model or the whole application? / I will send the evidence and name the unresolved question. / Does the interface display the response without changing it? / Let us inspect the boundary before assigning ownership. / Change one component while keeping the others fixed. / We have an observation, not a verified root cause. / Does this test depend on the policy index being current? / Please retain the configuration used for that request. / I can commit to an update, not to a fix by that time. / The demonstration remains scheduled; release approval is separate.
Vocabulary: application stack: The components that together deliver an application. / retrieval: Finding source material relevant to a request. / language model: A model that processes and generates language sequences. / user interface: The part of a system a user sees and operates. / orchestration: Coordination of model, retrieval, and tool steps. / inference: Running a trained model to produce an output. / endpoint: An address through which a service accepts requests. / payload: The data carried in a request or response. / trace: A linked record of events in a request's path. / dependency: A component or condition another part relies on. / upstream: Earlier in the relevant processing path. / downstream: Later in the relevant processing path. / symptom: An observed sign of a problem, not its cause. / root cause: The underlying cause established by investigation. / reproduction: Recreating a reported result under stated conditions. / request ID: An identifier used to locate a particular request. / schema: A defined structure for data and its fields. / tool call: A model-requested invocation of an available tool. / API: Application programming interface; a defined software interaction boundary. / contract: Agreed expectations between software components. / integration: Connecting components into a working system. / observability: The ability to infer system behavior from available signals. / triage: Initial assessment that determines priority and ownership. / handoff: Transfer of work and its relevant context.
Speaking task: Check the dialogue, then read the 30-day source and 60-day generated answer as separate observations. Locate the mismatch before assigning its cause. Switch roles. Ask for D-184, the redacted trace and the request configuration; retain the Thursday findings update without promising a fix. Complete the citation-link exchange. State that the generated response contains the citation while the interface omits it, then hand over the exact request.
Partner: Priya: AI integration engineer / Alex: Product manager
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A support assistant generates a correct answer with a citation ID. The interface omits the citation link. The display team needs the exact recorded request, not an assumption that retrieval failed.
Model for comparison AFTER my attempt, not a response to give me first: Alex: All right. I will describe the customer-visible problem and keep the demonstration decision separate. What should your update include for that decision?
Priya: The reproduction result, affected cases, remaining uncertainty, and a proposed next step. I will preserve the configuration so the comparison is meaningful.
Alex: That works. Please send it by three on Thursday, even if the investigation is incomplete. I will confirm who can decide about Friday.
Priya: Agreed. The handoff will say what we know, what we do not know, and which decision still needs an owner.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 01: Speaking the AI Development Stack
Original fictional case: A document assistant displays the wrong refund deadline. Retrieval returned the current policy, which states 30 days, but the generated answer says 60. The interface displays the generated text unchanged. Priya owns the model integration; Alex manages the product. A demonstration is scheduled for Friday. No cause has been confirmed, and Alex cannot approve a production release alone.
Language workshop: The answer is wrong, but which layer failed?
Communication goal: Distinguish components and ask a diagnostic question without assigning blame.
Language guidance: A trace records events; a diagnosis explains them. A wrong answer identifies a symptom, not automatically a model defect. Keep retrieval output, model output, and displayed text separate.
Useful expressions (preserve the supplied facts): Which component first shows the wrong value? / The retrieved passage says 30 days; the generated answer says 60. / That locates the mismatch, but it does not establish the cause. / Can you attach the request ID and the redacted trace? / Product approval and technical diagnosis are different decisions. / Can we reproduce it with the same payload and configuration? / By "the AI," do you mean the model or the whole application? / I will send the evidence and name the unresolved question. / Does the interface display the response without changing it? / Let us inspect the boundary before assigning ownership. / Change one component while keeping the others fixed. / We have an observation, not a verified root cause. / Does this test depend on the policy index being current? / Please retain the configuration used for that request. / I can commit to an update, not to a fix by that time. / The demonstration remains scheduled; release approval is separate.
Vocabulary: application stack: The components that together deliver an application. / retrieval: Finding source material relevant to a request. / language model: A model that processes and generates language sequences. / user interface: The part of a system a user sees and operates. / orchestration: Coordination of model, retrieval, and tool steps. / inference: Running a trained model to produce an output. / endpoint: An address through which a service accepts requests. / payload: The data carried in a request or response. / trace: A linked record of events in a request's path. / dependency: A component or condition another part relies on. / upstream: Earlier in the relevant processing path. / downstream: Later in the relevant processing path. / symptom: An observed sign of a problem, not its cause. / root cause: The underlying cause established by investigation. / reproduction: Recreating a reported result under stated conditions. / request ID: An identifier used to locate a particular request. / schema: A defined structure for data and its fields. / tool call: A model-requested invocation of an available tool. / API: Application programming interface; a defined software interaction boundary. / contract: Agreed expectations between software components. / integration: Connecting components into a working system. / observability: The ability to infer system behavior from available signals. / triage: Initial assessment that determines priority and ownership. / handoff: Transfer of work and its relevant context.
Speaking task: Check the dialogue, then read the 30-day source and 60-day generated answer as separate observations. Locate the mismatch before assigning its cause. Switch roles. Ask for D-184, the redacted trace and the request configuration; retain the Thursday findings update without promising a fix. Complete the citation-link exchange. State that the generated response contains the citation while the interface omits it, then hand over the exact request.
Partner: Priya: AI integration engineer / Alex: Product manager
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A support assistant generates a correct answer with a citation ID. The interface omits the citation link. The display team needs the exact recorded request, not an assumption that retrieval failed.
Model for comparison AFTER my attempt, not a response to give me first: Alex: All right. I will describe the customer-visible problem and keep the demonstration decision separate. What should your update include for that decision?
Priya: The reproduction result, affected cases, remaining uncertainty, and a proposed next step. I will preserve the configuration so the comparison is meaningful.
Alex: That works. Please send it by three on Thursday, even if the investigation is incomplete. I will confirm who can decide about Friday.
Priya: Agreed. The handoff will say what we know, what we do not know, and which decision still needs an owner.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 02: LLMs, Transformers, Tokens, and Context
Original fictional case: A document-summary request reserves 4,000 output tokens. Its instructions and documents contain 29,000 input tokens. The configured total context limit is 32,000 tokens, and the request is rejected before generation. Lina proposes increasing the output allowance. Omar owns the request configuration. The team must retain the exception clauses; no larger-context model has been evaluated for this task.
Language workshop: A larger answer will not make the input fit
Communication goal: Correct a technical misconception tactfully and negotiate a bounded alternative.
Language guidance: A token is not a word. Input size, reserved output, and total context are different quantities. The numerical limits in this case are fictional configuration facts, not specifications for a named model.
Useful expressions (preserve the supplied facts): You are proposing a larger output allowance, correct? / The request fails before generation because the combined reservation is too large. / Twenty-nine thousand plus four thousand is thirty-three thousand. / I see the aim, but that change would increase the reservation. / The allowance is a maximum, not the length already generated. / We must retain the exception clauses in the source context. / Could we remove exact duplicates before shortening substantive text? / A smaller output allowance is acceptable only if the summary still meets the task. / Are those counts from the tokenizer used by this model? / Fitting within the limit does not prove the answer is complete. / Which information would compression remove? / Please report truncation instead of silently dropping the end. / Let us compare the same documents under both configurations. / If the exceptions cannot fit, route the request for a different workflow. / The larger-context model needs its own quality check. / We will remove duplicates, recount, and test coverage before changing models.
Vocabulary: LLM: Large language model; a model that processes and generates language. / transformer: A neural architecture built around attention mechanisms. / token: A unit processed by a model, not necessarily a whole word. / tokenizer: The component that converts text into tokens. / context window: The amount of context a model can handle in a request. / input tokens: Tokens supplied to the model before generation. / output tokens: Tokens generated by the model. / token budget: An allocated number of tokens for a task or component. / system instruction: Higher-priority guidance supplied by an application's configuration. / prompt: Instructions and other context supplied for a model response. / attention: A mechanism that weights relationships between representations. / parameter: A learned numerical value within a model. / weight: A learned numerical coefficient in a model. / checkpoint: Saved model state from a particular stage. / truncation: Removal of content to fit a size limit. / compression: Reduction of information size, potentially losing detail. / reserved capacity: Capacity set aside rather than already consumed. / delimiter: A marker separating sections of supplied text. / instruction hierarchy: Priority relationships among different instruction sources. / long-context: Able to handle relatively large input context. / position: A token's place within the input sequence. / exception clause: Wording that limits or changes a general rule. / deduplication: Removal of repeated copies of content. / headroom: Unused capacity left below a limit.
Speaking task: After checking the dialogue, read the reservation calculation aloud: 29,000 input plus 4,000 reserved output equals 33,000 against a 32,000 limit. Switch roles. Read the proposal to remove exact duplicates while retaining the exception clauses; do not treat a fitting request as proof of coverage. Complete the appendix exchange. Check 18,000 plus 3,000 equals 21,000, then preserve unique obligations while comparing the total with the fictional 20,000 limit.
Partner: Lina: Application engineer / Omar: Model platform engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A fictional request has a 20,000-token total limit, 18,000 input tokens, and a 3,000-token output reservation. Repeated appendix covers may be removed; unique obligations must remain.
Model for comparison AFTER my attempt, not a response to give me first: Lina: I'll remove the repeated headers first. If we're still over, I'll test the smaller output allowance and bring both the counts and the coverage results.
Omar: Correct. Keep the delimiters between documents so the revised input does not merge separate policies into one undifferentiated block.
Lina: I will report the before-and-after counts with the coverage results. If the necessary clauses still cannot fit, I will stop rather than silently remove them.
Omar: Good. Describe the limit as a capacity constraint, not a model-quality finding. We can then choose the next workflow using the actual test results.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 02: LLMs, Transformers, Tokens, and Context
Original fictional case: A document-summary request reserves 4,000 output tokens. Its instructions and documents contain 29,000 input tokens. The configured total context limit is 32,000 tokens, and the request is rejected before generation. Lina proposes increasing the output allowance. Omar owns the request configuration. The team must retain the exception clauses; no larger-context model has been evaluated for this task.
Language workshop: A larger answer will not make the input fit
Communication goal: Correct a technical misconception tactfully and negotiate a bounded alternative.
Language guidance: A token is not a word. Input size, reserved output, and total context are different quantities. The numerical limits in this case are fictional configuration facts, not specifications for a named model.
Useful expressions (preserve the supplied facts): You are proposing a larger output allowance, correct? / The request fails before generation because the combined reservation is too large. / Twenty-nine thousand plus four thousand is thirty-three thousand. / I see the aim, but that change would increase the reservation. / The allowance is a maximum, not the length already generated. / We must retain the exception clauses in the source context. / Could we remove exact duplicates before shortening substantive text? / A smaller output allowance is acceptable only if the summary still meets the task. / Are those counts from the tokenizer used by this model? / Fitting within the limit does not prove the answer is complete. / Which information would compression remove? / Please report truncation instead of silently dropping the end. / Let us compare the same documents under both configurations. / If the exceptions cannot fit, route the request for a different workflow. / The larger-context model needs its own quality check. / We will remove duplicates, recount, and test coverage before changing models.
Vocabulary: LLM: Large language model; a model that processes and generates language. / transformer: A neural architecture built around attention mechanisms. / token: A unit processed by a model, not necessarily a whole word. / tokenizer: The component that converts text into tokens. / context window: The amount of context a model can handle in a request. / input tokens: Tokens supplied to the model before generation. / output tokens: Tokens generated by the model. / token budget: An allocated number of tokens for a task or component. / system instruction: Higher-priority guidance supplied by an application's configuration. / prompt: Instructions and other context supplied for a model response. / attention: A mechanism that weights relationships between representations. / parameter: A learned numerical value within a model. / weight: A learned numerical coefficient in a model. / checkpoint: Saved model state from a particular stage. / truncation: Removal of content to fit a size limit. / compression: Reduction of information size, potentially losing detail. / reserved capacity: Capacity set aside rather than already consumed. / delimiter: A marker separating sections of supplied text. / instruction hierarchy: Priority relationships among different instruction sources. / long-context: Able to handle relatively large input context. / position: A token's place within the input sequence. / exception clause: Wording that limits or changes a general rule. / deduplication: Removal of repeated copies of content. / headroom: Unused capacity left below a limit.
Speaking task: After checking the dialogue, read the reservation calculation aloud: 29,000 input plus 4,000 reserved output equals 33,000 against a 32,000 limit. Switch roles. Read the proposal to remove exact duplicates while retaining the exception clauses; do not treat a fitting request as proof of coverage. Complete the appendix exchange. Check 18,000 plus 3,000 equals 21,000, then preserve unique obligations while comparing the total with the fictional 20,000 limit.
Partner: Lina: Application engineer / Omar: Model platform engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A fictional request has a 20,000-token total limit, 18,000 input tokens, and a 3,000-token output reservation. Repeated appendix covers may be removed; unique obligations must remain.
Model for comparison AFTER my attempt, not a response to give me first: Lina: I'll remove the repeated headers first. If we're still over, I'll test the smaller output allowance and bring both the counts and the coverage results.
Omar: Correct. Keep the delimiters between documents so the revised input does not merge separate policies into one undifferentiated block.
Lina: I will report the before-and-after counts with the coverage results. If the necessary clauses still cannot fit, I will stop rather than silently remove them.
Omar: Good. Describe the limit as a capacity constraint, not a model-quality finding. We can then choose the next workflow using the actual test results.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 02: LLMs, Transformers, Tokens, and Context
Original fictional case: A document-summary request reserves 4,000 output tokens. Its instructions and documents contain 29,000 input tokens. The configured total context limit is 32,000 tokens, and the request is rejected before generation. Lina proposes increasing the output allowance. Omar owns the request configuration. The team must retain the exception clauses; no larger-context model has been evaluated for this task.
Language workshop: A larger answer will not make the input fit
Communication goal: Correct a technical misconception tactfully and negotiate a bounded alternative.
Language guidance: A token is not a word. Input size, reserved output, and total context are different quantities. The numerical limits in this case are fictional configuration facts, not specifications for a named model.
Useful expressions (preserve the supplied facts): You are proposing a larger output allowance, correct? / The request fails before generation because the combined reservation is too large. / Twenty-nine thousand plus four thousand is thirty-three thousand. / I see the aim, but that change would increase the reservation. / The allowance is a maximum, not the length already generated. / We must retain the exception clauses in the source context. / Could we remove exact duplicates before shortening substantive text? / A smaller output allowance is acceptable only if the summary still meets the task. / Are those counts from the tokenizer used by this model? / Fitting within the limit does not prove the answer is complete. / Which information would compression remove? / Please report truncation instead of silently dropping the end. / Let us compare the same documents under both configurations. / If the exceptions cannot fit, route the request for a different workflow. / The larger-context model needs its own quality check. / We will remove duplicates, recount, and test coverage before changing models.
Vocabulary: LLM: Large language model; a model that processes and generates language. / transformer: A neural architecture built around attention mechanisms. / token: A unit processed by a model, not necessarily a whole word. / tokenizer: The component that converts text into tokens. / context window: The amount of context a model can handle in a request. / input tokens: Tokens supplied to the model before generation. / output tokens: Tokens generated by the model. / token budget: An allocated number of tokens for a task or component. / system instruction: Higher-priority guidance supplied by an application's configuration. / prompt: Instructions and other context supplied for a model response. / attention: A mechanism that weights relationships between representations. / parameter: A learned numerical value within a model. / weight: A learned numerical coefficient in a model. / checkpoint: Saved model state from a particular stage. / truncation: Removal of content to fit a size limit. / compression: Reduction of information size, potentially losing detail. / reserved capacity: Capacity set aside rather than already consumed. / delimiter: A marker separating sections of supplied text. / instruction hierarchy: Priority relationships among different instruction sources. / long-context: Able to handle relatively large input context. / position: A token's place within the input sequence. / exception clause: Wording that limits or changes a general rule. / deduplication: Removal of repeated copies of content. / headroom: Unused capacity left below a limit.
Speaking task: After checking the dialogue, read the reservation calculation aloud: 29,000 input plus 4,000 reserved output equals 33,000 against a 32,000 limit. Switch roles. Read the proposal to remove exact duplicates while retaining the exception clauses; do not treat a fitting request as proof of coverage. Complete the appendix exchange. Check 18,000 plus 3,000 equals 21,000, then preserve unique obligations while comparing the total with the fictional 20,000 limit.
Partner: Lina: Application engineer / Omar: Model platform engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A fictional request has a 20,000-token total limit, 18,000 input tokens, and a 3,000-token output reservation. Repeated appendix covers may be removed; unique obligations must remain.
Model for comparison AFTER my attempt, not a response to give me first: Lina: I'll remove the repeated headers first. If we're still over, I'll test the smaller output allowance and bring both the counts and the coverage results.
Omar: Correct. Keep the delimiters between documents so the revised input does not merge separate policies into one undifferentiated block.
Lina: I will report the before-and-after counts with the coverage results. If the necessary clauses still cannot fit, I will stop rather than silently remove them.
Omar: Good. Describe the limit as a capacity constraint, not a model-quality finding. We can then choose the next workflow using the actual test results.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 03: Data, Datasets, Labels, and Leakage
Original fictional case: A classifier scores 94 percent on 500 evaluation examples. Jules finds that 40 evaluation records have exact copies in training. The remaining records have not been re-scored separately. Morgan presents release evidence in one hour. Training and evaluation labels use the same written rubric, but six disputed labels remain unresolved. The release owner can postpone the decision.
Language workshop: A 94 percent score with an overlap problem
Communication goal: Disclose compromised evidence and resist pressure to overstate readiness.
Language guidance: A clean holdout is separate from the process it evaluates, not necessarily representative of production. Removing overlap improves separation but does not resolve label disputes or distribution differences.
Useful expressions (preserve the supplied facts): The headline score includes records that overlap with training. / We found 40 exact overlaps among 500 evaluation examples. / We have not yet calculated the score on the remaining records. / Removing eight percent of records does not tell us the new accuracy. / Data overlap and disputed labels need different checks. / Can we trace when these records entered each split? / The current score should not be presented as an independent estimate. / I can provide the overlap list before the meeting. / What comparison rule defines an exact match here? / Do not tune against the replacement test set. / A clean split still needs a representativeness check. / Which label criterion produced the disagreement? / Please report both the number correct and the number evaluated. / This sample underrepresents the rare class. / Retain the original score with its limitation and the corrected result. / The release owner needs to decide whether to wait for the revised evidence.
Vocabulary: dataset: A collection of records assembled for a purpose. / training set: Data used to fit model parameters. / validation set: Data used to guide model selection or tuning. / test set: Data reserved for evaluating a finalized choice. / holdout: Data kept separate from a fitting or selection process. / data leakage: Information crossing a boundary in a way that invalidates evaluation. / contamination: Unwanted overlap or exposure that weakens an evaluation. / exact duplicate: A record identical under a stated comparison rule. / near duplicate: A closely similar record that may share substantive content. / split: A partition of data into separate subsets. / label: A target category or annotation attached to a record. / annotation: Human or machine-added information about an example. / rubric: Explicit criteria for assigning or evaluating labels. / adjudication: Resolution of a disputed judgment by a defined process. / inter-rater agreement: The degree to which different raters agree. / provenance: The recorded origin and history of data. / sampling frame: The population or list from which records are sampled. / representativeness: How well a sample reflects the target population. / class imbalance: Unequal frequencies among target categories. / stratification: Sampling or partitioning within defined groups. / distribution shift: A difference between data distributions across settings. / deduplicate: Remove repeated records under a defined rule. / denominator: The total count used as the base of a rate. / re-score: Calculate a score again under specified conditions.
Speaking task: Check the answers. Read 40 overlaps out of 500 as eight percent of records, not eight percentage points to subtract from the 94 percent score. Switch roles. Separate the overlap list, unmeasured clean-subset score and six disputed labels in the handoff. Complete the 80-message exchange. Retain 70 unscored nonoverlapping messages and three label disputes; do not invent a revised accuracy.
Partner: Jules: Data scientist / Morgan: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: An 80-message test contains ten copies of messages used for training. The team has not scored the 70 remaining messages separately. Two reviewers disagree on three labels.
Model for comparison AFTER my attempt, not a response to give me first: Morgan: All right, the duplicate check comes first. For the replacement set, can we keep it untouched while the team tunes the model?
Jules: It should. If we repeatedly choose model settings from that set, it becomes part of model selection, and we need a separate final test.
Morgan: I will recommend postponing the readiness decision. The slide will list confirmed overlap, pending label review, and the missing clean-subset score separately.
Jules: I will send the records and a short limitation statement. That gives the owner usable evidence without pretending the corrected outcome is already known.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 03: Data, Datasets, Labels, and Leakage
Original fictional case: A classifier scores 94 percent on 500 evaluation examples. Jules finds that 40 evaluation records have exact copies in training. The remaining records have not been re-scored separately. Morgan presents release evidence in one hour. Training and evaluation labels use the same written rubric, but six disputed labels remain unresolved. The release owner can postpone the decision.
Language workshop: A 94 percent score with an overlap problem
Communication goal: Disclose compromised evidence and resist pressure to overstate readiness.
Language guidance: A clean holdout is separate from the process it evaluates, not necessarily representative of production. Removing overlap improves separation but does not resolve label disputes or distribution differences.
Useful expressions (preserve the supplied facts): The headline score includes records that overlap with training. / We found 40 exact overlaps among 500 evaluation examples. / We have not yet calculated the score on the remaining records. / Removing eight percent of records does not tell us the new accuracy. / Data overlap and disputed labels need different checks. / Can we trace when these records entered each split? / The current score should not be presented as an independent estimate. / I can provide the overlap list before the meeting. / What comparison rule defines an exact match here? / Do not tune against the replacement test set. / A clean split still needs a representativeness check. / Which label criterion produced the disagreement? / Please report both the number correct and the number evaluated. / This sample underrepresents the rare class. / Retain the original score with its limitation and the corrected result. / The release owner needs to decide whether to wait for the revised evidence.
Vocabulary: dataset: A collection of records assembled for a purpose. / training set: Data used to fit model parameters. / validation set: Data used to guide model selection or tuning. / test set: Data reserved for evaluating a finalized choice. / holdout: Data kept separate from a fitting or selection process. / data leakage: Information crossing a boundary in a way that invalidates evaluation. / contamination: Unwanted overlap or exposure that weakens an evaluation. / exact duplicate: A record identical under a stated comparison rule. / near duplicate: A closely similar record that may share substantive content. / split: A partition of data into separate subsets. / label: A target category or annotation attached to a record. / annotation: Human or machine-added information about an example. / rubric: Explicit criteria for assigning or evaluating labels. / adjudication: Resolution of a disputed judgment by a defined process. / inter-rater agreement: The degree to which different raters agree. / provenance: The recorded origin and history of data. / sampling frame: The population or list from which records are sampled. / representativeness: How well a sample reflects the target population. / class imbalance: Unequal frequencies among target categories. / stratification: Sampling or partitioning within defined groups. / distribution shift: A difference between data distributions across settings. / deduplicate: Remove repeated records under a defined rule. / denominator: The total count used as the base of a rate. / re-score: Calculate a score again under specified conditions.
Speaking task: Check the answers. Read 40 overlaps out of 500 as eight percent of records, not eight percentage points to subtract from the 94 percent score. Switch roles. Separate the overlap list, unmeasured clean-subset score and six disputed labels in the handoff. Complete the 80-message exchange. Retain 70 unscored nonoverlapping messages and three label disputes; do not invent a revised accuracy.
Partner: Jules: Data scientist / Morgan: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: An 80-message test contains ten copies of messages used for training. The team has not scored the 70 remaining messages separately. Two reviewers disagree on three labels.
Model for comparison AFTER my attempt, not a response to give me first: Morgan: All right, the duplicate check comes first. For the replacement set, can we keep it untouched while the team tunes the model?
Jules: It should. If we repeatedly choose model settings from that set, it becomes part of model selection, and we need a separate final test.
Morgan: I will recommend postponing the readiness decision. The slide will list confirmed overlap, pending label review, and the missing clean-subset score separately.
Jules: I will send the records and a short limitation statement. That gives the owner usable evidence without pretending the corrected outcome is already known.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 03: Data, Datasets, Labels, and Leakage
Original fictional case: A classifier scores 94 percent on 500 evaluation examples. Jules finds that 40 evaluation records have exact copies in training. The remaining records have not been re-scored separately. Morgan presents release evidence in one hour. Training and evaluation labels use the same written rubric, but six disputed labels remain unresolved. The release owner can postpone the decision.
Language workshop: A 94 percent score with an overlap problem
Communication goal: Disclose compromised evidence and resist pressure to overstate readiness.
Language guidance: A clean holdout is separate from the process it evaluates, not necessarily representative of production. Removing overlap improves separation but does not resolve label disputes or distribution differences.
Useful expressions (preserve the supplied facts): The headline score includes records that overlap with training. / We found 40 exact overlaps among 500 evaluation examples. / We have not yet calculated the score on the remaining records. / Removing eight percent of records does not tell us the new accuracy. / Data overlap and disputed labels need different checks. / Can we trace when these records entered each split? / The current score should not be presented as an independent estimate. / I can provide the overlap list before the meeting. / What comparison rule defines an exact match here? / Do not tune against the replacement test set. / A clean split still needs a representativeness check. / Which label criterion produced the disagreement? / Please report both the number correct and the number evaluated. / This sample underrepresents the rare class. / Retain the original score with its limitation and the corrected result. / The release owner needs to decide whether to wait for the revised evidence.
Vocabulary: dataset: A collection of records assembled for a purpose. / training set: Data used to fit model parameters. / validation set: Data used to guide model selection or tuning. / test set: Data reserved for evaluating a finalized choice. / holdout: Data kept separate from a fitting or selection process. / data leakage: Information crossing a boundary in a way that invalidates evaluation. / contamination: Unwanted overlap or exposure that weakens an evaluation. / exact duplicate: A record identical under a stated comparison rule. / near duplicate: A closely similar record that may share substantive content. / split: A partition of data into separate subsets. / label: A target category or annotation attached to a record. / annotation: Human or machine-added information about an example. / rubric: Explicit criteria for assigning or evaluating labels. / adjudication: Resolution of a disputed judgment by a defined process. / inter-rater agreement: The degree to which different raters agree. / provenance: The recorded origin and history of data. / sampling frame: The population or list from which records are sampled. / representativeness: How well a sample reflects the target population. / class imbalance: Unequal frequencies among target categories. / stratification: Sampling or partitioning within defined groups. / distribution shift: A difference between data distributions across settings. / deduplicate: Remove repeated records under a defined rule. / denominator: The total count used as the base of a rate. / re-score: Calculate a score again under specified conditions.
Speaking task: Check the answers. Read 40 overlaps out of 500 as eight percent of records, not eight percentage points to subtract from the 94 percent score. Switch roles. Separate the overlap list, unmeasured clean-subset score and six disputed labels in the handoff. Complete the 80-message exchange. Retain 70 unscored nonoverlapping messages and three label disputes; do not invent a revised accuracy.
Partner: Jules: Data scientist / Morgan: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: An 80-message test contains ten copies of messages used for training. The team has not scored the 70 remaining messages separately. Two reviewers disagree on three labels.
Model for comparison AFTER my attempt, not a response to give me first: Morgan: All right, the duplicate check comes first. For the replacement set, can we keep it untouched while the team tunes the model?
Jules: It should. If we repeatedly choose model settings from that set, it becomes part of model selection, and we need a separate final test.
Morgan: I will recommend postponing the readiness decision. The slide will list confirmed overlap, pending label review, and the missing clean-subset score separately.
Jules: I will send the records and a short limitation statement. That gives the owner usable evidence without pretending the corrected outcome is already known.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 04: Retrieval, Embeddings, Vector Search, and RAG
Original fictional case: A document assistant with reranking retrieves a relevant passage for 86 of 100 fixed test queries, compared with 78 without reranking. The reranker adds 300 milliseconds to the measured retrieval path. The generator and query set are unchanged. No user study or end-to-end answer-quality review has been completed. Sam recommends a pilot; Elena owns the evaluation plan.
Language workshop: More relevant passages, a slower response
Communication goal: Negotiate a trade-off using matched comparisons and explicitly limited claims.
Language guidance: Retrieval relevance, answer correctness, and user preference are distinct outcomes. A citation can identify a source without proving that the source supports the exact claim. Inspect the cited passage.
Useful expressions (preserve the supplied facts): The improvement is passage coverage on this fixed query set. / We have not yet measured final-answer accuracy. / The gain comes with 300 milliseconds of added retrieval time. / Can we hold the generator and queries constant? / What happens when we remove only the reranker? / These results support a bounded pilot, not a general rollout claim. / Does the cited passage support this specific sentence? / Review final answers and timing on the same cases. / Was the relevant passage absent or merely ranked too low? / Did a filter exclude the document before similarity search? / The exception and the main rule should not be separated without context. / Which query types account for most of the added time? / A relevant passage may still come from an outdated index. / The matched test isolates this change, not every production condition. / Let us compare reranking all queries with a defined routing policy. / User preference remains unmeasured in this experiment.
Vocabulary: RAG: Retrieval-augmented generation; generating with retrieved context. / embedding: A numerical representation used to compare or process items. / vector: An ordered collection of numerical values. / vector store: A store or index supporting retrieval of vector representations. / similarity search: Finding items close under a chosen similarity measure. / semantic search: Retrieval based on meaning-related representations. / keyword search: Retrieval based on words or lexical matches. / hybrid retrieval: Retrieval combining different search approaches. / chunk: A retrievable segment of a larger document. / chunk boundary: The point where one segment ends and another starts. / overlap window: Content shared between adjacent chunks. / reranker: A component that reorders candidates for relevance. / candidate set: Items available for a later selection or ranking step. / top-k: The first k results returned by a ranking. / relevance: How well material addresses a specified query. / recall: The proportion of relevant items retrieved under a defined evaluation. / precision: The proportion of retrieved items judged relevant. / grounding: Connecting generated claims to supporting information. / citation: A reference identifying a supporting source. / source attribution: Identification of where information came from. / index freshness: How current the indexed information is. / filter: A rule limiting eligible retrieval results. / ablation: A comparison that removes or changes a component. / end-to-end: Covering the complete path from request to outcome.
Speaking task: Check the dialogue. Read 78 versus 86 of 100: eight percentage points gained, with 300 milliseconds added. Switch roles. Request a matched final-answer review; do not substitute retrieval coverage for answer accuracy. Complete the exception exchange. Distinguish a real source from support for the specific claim.
Partner: Sam: Retrieval engineer / Elena: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A benefits assistant cites the current policy but uses the standard rule instead of a stated exception. Retrieval includes both sections. A reviewer must report claim support, not count citation markers.
Model for comparison AFTER my attempt, not a response to give me first: Sam: Then I will present the retrieval result as eight percentage points on this set. I will not call it an eight-percent increase in customer satisfaction.
Elena: Exactly. Also break down timing by query type, because the same average can hide very different costs for simple and complex questions.
Sam: I'll bring the answer review and timing breakdown before we discuss the pilot. The chunking experiment can wait; I don't want to lose this comparison.
Elena: Good. Put the improvement on the slide, but keep the pilot recommendation conditional. I want to see the actual answers before I approve it.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 04: Retrieval, Embeddings, Vector Search, and RAG
Original fictional case: A document assistant with reranking retrieves a relevant passage for 86 of 100 fixed test queries, compared with 78 without reranking. The reranker adds 300 milliseconds to the measured retrieval path. The generator and query set are unchanged. No user study or end-to-end answer-quality review has been completed. Sam recommends a pilot; Elena owns the evaluation plan.
Language workshop: More relevant passages, a slower response
Communication goal: Negotiate a trade-off using matched comparisons and explicitly limited claims.
Language guidance: Retrieval relevance, answer correctness, and user preference are distinct outcomes. A citation can identify a source without proving that the source supports the exact claim. Inspect the cited passage.
Useful expressions (preserve the supplied facts): The improvement is passage coverage on this fixed query set. / We have not yet measured final-answer accuracy. / The gain comes with 300 milliseconds of added retrieval time. / Can we hold the generator and queries constant? / What happens when we remove only the reranker? / These results support a bounded pilot, not a general rollout claim. / Does the cited passage support this specific sentence? / Review final answers and timing on the same cases. / Was the relevant passage absent or merely ranked too low? / Did a filter exclude the document before similarity search? / The exception and the main rule should not be separated without context. / Which query types account for most of the added time? / A relevant passage may still come from an outdated index. / The matched test isolates this change, not every production condition. / Let us compare reranking all queries with a defined routing policy. / User preference remains unmeasured in this experiment.
Vocabulary: RAG: Retrieval-augmented generation; generating with retrieved context. / embedding: A numerical representation used to compare or process items. / vector: An ordered collection of numerical values. / vector store: A store or index supporting retrieval of vector representations. / similarity search: Finding items close under a chosen similarity measure. / semantic search: Retrieval based on meaning-related representations. / keyword search: Retrieval based on words or lexical matches. / hybrid retrieval: Retrieval combining different search approaches. / chunk: A retrievable segment of a larger document. / chunk boundary: The point where one segment ends and another starts. / overlap window: Content shared between adjacent chunks. / reranker: A component that reorders candidates for relevance. / candidate set: Items available for a later selection or ranking step. / top-k: The first k results returned by a ranking. / relevance: How well material addresses a specified query. / recall: The proportion of relevant items retrieved under a defined evaluation. / precision: The proportion of retrieved items judged relevant. / grounding: Connecting generated claims to supporting information. / citation: A reference identifying a supporting source. / source attribution: Identification of where information came from. / index freshness: How current the indexed information is. / filter: A rule limiting eligible retrieval results. / ablation: A comparison that removes or changes a component. / end-to-end: Covering the complete path from request to outcome.
Speaking task: Check the dialogue. Read 78 versus 86 of 100: eight percentage points gained, with 300 milliseconds added. Switch roles. Request a matched final-answer review; do not substitute retrieval coverage for answer accuracy. Complete the exception exchange. Distinguish a real source from support for the specific claim.
Partner: Sam: Retrieval engineer / Elena: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A benefits assistant cites the current policy but uses the standard rule instead of a stated exception. Retrieval includes both sections. A reviewer must report claim support, not count citation markers.
Model for comparison AFTER my attempt, not a response to give me first: Sam: Then I will present the retrieval result as eight percentage points on this set. I will not call it an eight-percent increase in customer satisfaction.
Elena: Exactly. Also break down timing by query type, because the same average can hide very different costs for simple and complex questions.
Sam: I'll bring the answer review and timing breakdown before we discuss the pilot. The chunking experiment can wait; I don't want to lose this comparison.
Elena: Good. Put the improvement on the slide, but keep the pilot recommendation conditional. I want to see the actual answers before I approve it.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 04: Retrieval, Embeddings, Vector Search, and RAG
Original fictional case: A document assistant with reranking retrieves a relevant passage for 86 of 100 fixed test queries, compared with 78 without reranking. The reranker adds 300 milliseconds to the measured retrieval path. The generator and query set are unchanged. No user study or end-to-end answer-quality review has been completed. Sam recommends a pilot; Elena owns the evaluation plan.
Language workshop: More relevant passages, a slower response
Communication goal: Negotiate a trade-off using matched comparisons and explicitly limited claims.
Language guidance: Retrieval relevance, answer correctness, and user preference are distinct outcomes. A citation can identify a source without proving that the source supports the exact claim. Inspect the cited passage.
Useful expressions (preserve the supplied facts): The improvement is passage coverage on this fixed query set. / We have not yet measured final-answer accuracy. / The gain comes with 300 milliseconds of added retrieval time. / Can we hold the generator and queries constant? / What happens when we remove only the reranker? / These results support a bounded pilot, not a general rollout claim. / Does the cited passage support this specific sentence? / Review final answers and timing on the same cases. / Was the relevant passage absent or merely ranked too low? / Did a filter exclude the document before similarity search? / The exception and the main rule should not be separated without context. / Which query types account for most of the added time? / A relevant passage may still come from an outdated index. / The matched test isolates this change, not every production condition. / Let us compare reranking all queries with a defined routing policy. / User preference remains unmeasured in this experiment.
Vocabulary: RAG: Retrieval-augmented generation; generating with retrieved context. / embedding: A numerical representation used to compare or process items. / vector: An ordered collection of numerical values. / vector store: A store or index supporting retrieval of vector representations. / similarity search: Finding items close under a chosen similarity measure. / semantic search: Retrieval based on meaning-related representations. / keyword search: Retrieval based on words or lexical matches. / hybrid retrieval: Retrieval combining different search approaches. / chunk: A retrievable segment of a larger document. / chunk boundary: The point where one segment ends and another starts. / overlap window: Content shared between adjacent chunks. / reranker: A component that reorders candidates for relevance. / candidate set: Items available for a later selection or ranking step. / top-k: The first k results returned by a ranking. / relevance: How well material addresses a specified query. / recall: The proportion of relevant items retrieved under a defined evaluation. / precision: The proportion of retrieved items judged relevant. / grounding: Connecting generated claims to supporting information. / citation: A reference identifying a supporting source. / source attribution: Identification of where information came from. / index freshness: How current the indexed information is. / filter: A rule limiting eligible retrieval results. / ablation: A comparison that removes or changes a component. / end-to-end: Covering the complete path from request to outcome.
Speaking task: Check the dialogue. Read 78 versus 86 of 100: eight percentage points gained, with 300 milliseconds added. Switch roles. Request a matched final-answer review; do not substitute retrieval coverage for answer accuracy. Complete the exception exchange. Distinguish a real source from support for the specific claim.
Partner: Sam: Retrieval engineer / Elena: Evaluation lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A benefits assistant cites the current policy but uses the standard rule instead of a stated exception. Retrieval includes both sections. A reviewer must report claim support, not count citation markers.
Model for comparison AFTER my attempt, not a response to give me first: Sam: Then I will present the retrieval result as eight percentage points on this set. I will not call it an eight-percent increase in customer satisfaction.
Elena: Exactly. Also break down timing by query type, because the same average can hide very different costs for simple and complex questions.
Sam: I'll bring the answer review and timing breakdown before we discuss the pilot. The chunking experiment can wait; I don't want to lose this comparison.
Elena: Good. Put the improvement on the slide, but keep the pilot recommendation conditional. I want to see the actual answers before I approve it.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 05: Fine-Tuning, Alignment, and Adaptation
Original fictional case: A support summarizer sometimes returns paragraphs instead of the required fields and sometimes cites an obsolete refund policy. The obsolete wording is present in its retrieval index. Its prompt contains only one formatting example. Tessa proposes fine-tuning to solve both problems. Dev owns the experiment plan. The team has 200 permission-cleared example summaries but has not audited their label consistency.
Language workshop: One proposal, two different failure modes
Communication goal: Separate proposed remedies by mechanism and agree on a controlled experiment.
Language guidance: Fine-tuning changes model parameters; prompting supplies instructions at inference time; retrieval supplies external context. These interventions can interact, but they do not have interchangeable mechanisms or guarantees.
Useful expressions (preserve the supplied facts): We have a formatting issue and a source-freshness issue. / How would this change prevent the obsolete policy from entering the context? / The target is consistent output in the required fields. / Let us test clearer prompting before attributing gains to training. / Are the reference summaries consistent with the same rubric? / Fine-tuning may improve the behavior; it does not keep an index current. / Use the same cleaned source set for both configurations. / We will compare format adherence and factual correctness separately. / Keep the final test examples outside the training set. / A constrained output format may address part of this problem. / Are these examples cleared for the intended training use? / Permission, quality, and representativeness are separate checks. / Does the adaptation reduce performance on other required tasks? / Retain the validated base configuration for comparison. / I recommend an experiment, not immediate adoption. / We will fix the stale index and test the formatting intervention independently.
Vocabulary: fine-tuning: Updating model parameters using additional training data. / pretraining: Broad initial training before task-specific adaptation. / supervised fine-tuning: Training on supplied input-output examples. / SFT: Abbreviation for supervised fine-tuning. / preference data: Examples recording preferences between candidate outputs. / alignment: Shaping behavior toward specified goals or preferences. / RLHF: Reinforcement learning from human feedback. / DPO: Direct preference optimization; a preference-training method. / LoRA: Low-rank adaptation; a parameter-efficient adaptation method. / adapter: A small trainable component used for adaptation. / base model: The starting model before the adaptation under discussion. / training objective: The quantity a training process seeks to optimize. / hyperparameter: A training or model setting selected rather than learned in that run. / learning rate: A setting controlling the scale of training updates. / epoch: One pass through the training data. / overfitting: Fitting training patterns without adequate generalization. / generalization: Performance on relevant examples beyond those used for fitting. / catastrophic forgetting: Loss of previously learned capabilities during further training. / format adherence: Consistency with a required output structure. / structured output: Output constrained or organized into a defined structure. / few-shot prompting: Supplying a small set of examples in the prompt. / behavioral target: The specified behavior an intervention aims to change. / training artifact: A saved output of training, such as weights or configuration. / rollback candidate: A previously validated configuration available for reversion.
Speaking task: After checking the answers, identify the two failures aloud: obsolete source wording and inconsistent output fields. Switch roles. Read the proposed experiment with fixed sources and test cases; keep data-use permission separate from the unfinished label audit. Complete the category-list exchange. Correct the six-category prompt to the stated five-category requirement, while leaving any training benefit unconfirmed.
Partner: Tessa: Applied scientist / Dev: Technical lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier is required to use five current categories. Its prompt still lists six older categories. The team has not yet tested a corrected prompt or new training data.
Model for comparison AFTER my attempt, not a response to give me first: Tessa: I will version the examples, configuration, and resulting adapter together. That way the team can reproduce the comparison and identify exactly what was evaluated.
Dev: Good. Retain a validated rollback candidate as well. A successful experiment is not a reason to lose the previous working configuration.
Tessa: I'll start with the example audit and the prompting baseline. Can we review those results before we spend time on an adapter?
Dev: Yes. List the expected gains as hypotheses in the experiment note. If the baseline solves the problem, we may not need training.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 05: Fine-Tuning, Alignment, and Adaptation
Original fictional case: A support summarizer sometimes returns paragraphs instead of the required fields and sometimes cites an obsolete refund policy. The obsolete wording is present in its retrieval index. Its prompt contains only one formatting example. Tessa proposes fine-tuning to solve both problems. Dev owns the experiment plan. The team has 200 permission-cleared example summaries but has not audited their label consistency.
Language workshop: One proposal, two different failure modes
Communication goal: Separate proposed remedies by mechanism and agree on a controlled experiment.
Language guidance: Fine-tuning changes model parameters; prompting supplies instructions at inference time; retrieval supplies external context. These interventions can interact, but they do not have interchangeable mechanisms or guarantees.
Useful expressions (preserve the supplied facts): We have a formatting issue and a source-freshness issue. / How would this change prevent the obsolete policy from entering the context? / The target is consistent output in the required fields. / Let us test clearer prompting before attributing gains to training. / Are the reference summaries consistent with the same rubric? / Fine-tuning may improve the behavior; it does not keep an index current. / Use the same cleaned source set for both configurations. / We will compare format adherence and factual correctness separately. / Keep the final test examples outside the training set. / A constrained output format may address part of this problem. / Are these examples cleared for the intended training use? / Permission, quality, and representativeness are separate checks. / Does the adaptation reduce performance on other required tasks? / Retain the validated base configuration for comparison. / I recommend an experiment, not immediate adoption. / We will fix the stale index and test the formatting intervention independently.
Vocabulary: fine-tuning: Updating model parameters using additional training data. / pretraining: Broad initial training before task-specific adaptation. / supervised fine-tuning: Training on supplied input-output examples. / SFT: Abbreviation for supervised fine-tuning. / preference data: Examples recording preferences between candidate outputs. / alignment: Shaping behavior toward specified goals or preferences. / RLHF: Reinforcement learning from human feedback. / DPO: Direct preference optimization; a preference-training method. / LoRA: Low-rank adaptation; a parameter-efficient adaptation method. / adapter: A small trainable component used for adaptation. / base model: The starting model before the adaptation under discussion. / training objective: The quantity a training process seeks to optimize. / hyperparameter: A training or model setting selected rather than learned in that run. / learning rate: A setting controlling the scale of training updates. / epoch: One pass through the training data. / overfitting: Fitting training patterns without adequate generalization. / generalization: Performance on relevant examples beyond those used for fitting. / catastrophic forgetting: Loss of previously learned capabilities during further training. / format adherence: Consistency with a required output structure. / structured output: Output constrained or organized into a defined structure. / few-shot prompting: Supplying a small set of examples in the prompt. / behavioral target: The specified behavior an intervention aims to change. / training artifact: A saved output of training, such as weights or configuration. / rollback candidate: A previously validated configuration available for reversion.
Speaking task: After checking the answers, identify the two failures aloud: obsolete source wording and inconsistent output fields. Switch roles. Read the proposed experiment with fixed sources and test cases; keep data-use permission separate from the unfinished label audit. Complete the category-list exchange. Correct the six-category prompt to the stated five-category requirement, while leaving any training benefit unconfirmed.
Partner: Tessa: Applied scientist / Dev: Technical lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier is required to use five current categories. Its prompt still lists six older categories. The team has not yet tested a corrected prompt or new training data.
Model for comparison AFTER my attempt, not a response to give me first: Tessa: I will version the examples, configuration, and resulting adapter together. That way the team can reproduce the comparison and identify exactly what was evaluated.
Dev: Good. Retain a validated rollback candidate as well. A successful experiment is not a reason to lose the previous working configuration.
Tessa: I'll start with the example audit and the prompting baseline. Can we review those results before we spend time on an adapter?
Dev: Yes. List the expected gains as hypotheses in the experiment note. If the baseline solves the problem, we may not need training.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 05: Fine-Tuning, Alignment, and Adaptation
Original fictional case: A support summarizer sometimes returns paragraphs instead of the required fields and sometimes cites an obsolete refund policy. The obsolete wording is present in its retrieval index. Its prompt contains only one formatting example. Tessa proposes fine-tuning to solve both problems. Dev owns the experiment plan. The team has 200 permission-cleared example summaries but has not audited their label consistency.
Language workshop: One proposal, two different failure modes
Communication goal: Separate proposed remedies by mechanism and agree on a controlled experiment.
Language guidance: Fine-tuning changes model parameters; prompting supplies instructions at inference time; retrieval supplies external context. These interventions can interact, but they do not have interchangeable mechanisms or guarantees.
Useful expressions (preserve the supplied facts): We have a formatting issue and a source-freshness issue. / How would this change prevent the obsolete policy from entering the context? / The target is consistent output in the required fields. / Let us test clearer prompting before attributing gains to training. / Are the reference summaries consistent with the same rubric? / Fine-tuning may improve the behavior; it does not keep an index current. / Use the same cleaned source set for both configurations. / We will compare format adherence and factual correctness separately. / Keep the final test examples outside the training set. / A constrained output format may address part of this problem. / Are these examples cleared for the intended training use? / Permission, quality, and representativeness are separate checks. / Does the adaptation reduce performance on other required tasks? / Retain the validated base configuration for comparison. / I recommend an experiment, not immediate adoption. / We will fix the stale index and test the formatting intervention independently.
Vocabulary: fine-tuning: Updating model parameters using additional training data. / pretraining: Broad initial training before task-specific adaptation. / supervised fine-tuning: Training on supplied input-output examples. / SFT: Abbreviation for supervised fine-tuning. / preference data: Examples recording preferences between candidate outputs. / alignment: Shaping behavior toward specified goals or preferences. / RLHF: Reinforcement learning from human feedback. / DPO: Direct preference optimization; a preference-training method. / LoRA: Low-rank adaptation; a parameter-efficient adaptation method. / adapter: A small trainable component used for adaptation. / base model: The starting model before the adaptation under discussion. / training objective: The quantity a training process seeks to optimize. / hyperparameter: A training or model setting selected rather than learned in that run. / learning rate: A setting controlling the scale of training updates. / epoch: One pass through the training data. / overfitting: Fitting training patterns without adequate generalization. / generalization: Performance on relevant examples beyond those used for fitting. / catastrophic forgetting: Loss of previously learned capabilities during further training. / format adherence: Consistency with a required output structure. / structured output: Output constrained or organized into a defined structure. / few-shot prompting: Supplying a small set of examples in the prompt. / behavioral target: The specified behavior an intervention aims to change. / training artifact: A saved output of training, such as weights or configuration. / rollback candidate: A previously validated configuration available for reversion.
Speaking task: After checking the answers, identify the two failures aloud: obsolete source wording and inconsistent output fields. Switch roles. Read the proposed experiment with fixed sources and test cases; keep data-use permission separate from the unfinished label audit. Complete the category-list exchange. Correct the six-category prompt to the stated five-category requirement, while leaving any training benefit unconfirmed.
Partner: Tessa: Applied scientist / Dev: Technical lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier is required to use five current categories. Its prompt still lists six older categories. The team has not yet tested a corrected prompt or new training data.
Model for comparison AFTER my attempt, not a response to give me first: Tessa: I will version the examples, configuration, and resulting adapter together. That way the team can reproduce the comparison and identify exactly what was evaluated.
Dev: Good. Retain a validated rollback candidate as well. A successful experiment is not a reason to lose the previous working configuration.
Tessa: I'll start with the example audit and the prompting baseline. Can we review those results before we spend time on an adapter?
Dev: Yes. List the expected gains as hypotheses in the experiment note. If the baseline solves the problem, we may not need training.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 06: Evaluation, Benchmarks, and Regression
Original fictional case: On the same 100-case test, the previous model passes 90 cases and the candidate passes 88. Among 20 long-document cases, passes rise from 12 to 18. Among 80 short-request cases, passes fall from 78 to 70. The scoring rubric is unchanged. Production traffic proportions and the seriousness of individual failures have not been analyzed. Noor chairs the review; Ben prepared the results.
Language workshop: The average is down, but one slice improved
Communication goal: Present disaggregated results and challenge an oversimplified release argument.
Language guidance: The aggregate falls by two percentage points. Long-document performance rises from 60 to 90 percent; short-request performance falls from 97.5 to 87.5 percent. These are test-set rates, not verified production estimates.
Useful expressions (preserve the supplied facts): Overall passes fell, while the long-document slice improved. / Eighteen of 20 long-document cases passed. / Short-request passes fell from 78 to 70 out of 80. / Which previously passing cases now fail? / A count alone does not tell us how serious those failures are. / Does this test mixture resemble the traffic we plan to serve? / We cannot infer production performance from these counts alone. / Compare the changed cases and classify their failure severity. / The improved slice should not replace the overall result on the slide. / Were both versions scored against the same rubric? / We have not established whether this difference is stable across samples. / Which acceptance criterion applies to the short-request regression? / Let us agree on the criteria before selecting the winning score. / How closely does the automated judge match the reviewed labels? / Routing long documents separately is a proposal to test, not an approved fix. / We will defer the decision until the changed cases have been assessed.
Vocabulary: evaluation: A structured assessment against specified criteria. / benchmark: A reference test used for comparison. / baseline: The reference system or result for a comparison. / candidate: The version under consideration. / regression: A deterioration in previously measured behavior. / test case: A defined input and expected criterion or outcome. / golden set: Curated cases used to check important behavior repeatedly. / pass rate: The fraction of evaluated cases meeting a criterion. / acceptance criterion: A stated condition required for acceptance. / slice: A defined subgroup within evaluation data. / aggregate: A combined measure across multiple cases or groups. / disaggregation: Breaking a combined result into meaningful groups. / severity: The seriousness of a failure's consequences. / false positive: A positive decision when the reference outcome is negative. / false negative: A negative decision when the reference outcome is positive. / confusion matrix: A table comparing predicted and reference categories. / calibration: Agreement between stated confidence and observed outcomes. / confidence interval: A range expressing uncertainty under stated statistical assumptions. / statistical significance: Evidence against a null model under a specified statistical test. / practical significance: The importance of a difference for the actual task. / paired comparison: Comparison of results on corresponding cases. / LLM-as-judge: Use of a language model to evaluate outputs. / human review: Assessment by people using defined criteria. / release gate: A defined decision checkpoint before release.
Speaking task: Check the dialogue. Read both slices: 12 to 18 of 20 long cases; 78 to 70 of 80 short cases. Switch roles. State that total passes fall from 90 to 88; severity and production mixture remain unmeasured. Complete the urgent-message exchange. Name the four misses as false negatives and retain the required severity review.
Partner: Noor: Evaluation manager / Ben: Machine-learning engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier has a high overall score but misses four of ten urgent messages. Urgent-message consequences have not been reviewed. The team must not treat aggregate success as sufficient evidence for release.
Model for comparison AFTER my attempt, not a response to give me first: Ben: I will revise the slide to show both counts, the shared rubric, and the missing traffic and severity analysis. The candidate's strengths will still be visible.
Noor: Good. That is disaggregation, not an attempt to hide the average. We want the decision-maker to see why the headline alone is insufficient.
Ben: I'll recommend waiting for the changed-case review. I still think routing is worth testing, but I won't present it as a fix today.
Noor: Agreed. The release gate remains open. The next update should say which criteria are met, which are not, and what evidence is still missing.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 06: Evaluation, Benchmarks, and Regression
Original fictional case: On the same 100-case test, the previous model passes 90 cases and the candidate passes 88. Among 20 long-document cases, passes rise from 12 to 18. Among 80 short-request cases, passes fall from 78 to 70. The scoring rubric is unchanged. Production traffic proportions and the seriousness of individual failures have not been analyzed. Noor chairs the review; Ben prepared the results.
Language workshop: The average is down, but one slice improved
Communication goal: Present disaggregated results and challenge an oversimplified release argument.
Language guidance: The aggregate falls by two percentage points. Long-document performance rises from 60 to 90 percent; short-request performance falls from 97.5 to 87.5 percent. These are test-set rates, not verified production estimates.
Useful expressions (preserve the supplied facts): Overall passes fell, while the long-document slice improved. / Eighteen of 20 long-document cases passed. / Short-request passes fell from 78 to 70 out of 80. / Which previously passing cases now fail? / A count alone does not tell us how serious those failures are. / Does this test mixture resemble the traffic we plan to serve? / We cannot infer production performance from these counts alone. / Compare the changed cases and classify their failure severity. / The improved slice should not replace the overall result on the slide. / Were both versions scored against the same rubric? / We have not established whether this difference is stable across samples. / Which acceptance criterion applies to the short-request regression? / Let us agree on the criteria before selecting the winning score. / How closely does the automated judge match the reviewed labels? / Routing long documents separately is a proposal to test, not an approved fix. / We will defer the decision until the changed cases have been assessed.
Vocabulary: evaluation: A structured assessment against specified criteria. / benchmark: A reference test used for comparison. / baseline: The reference system or result for a comparison. / candidate: The version under consideration. / regression: A deterioration in previously measured behavior. / test case: A defined input and expected criterion or outcome. / golden set: Curated cases used to check important behavior repeatedly. / pass rate: The fraction of evaluated cases meeting a criterion. / acceptance criterion: A stated condition required for acceptance. / slice: A defined subgroup within evaluation data. / aggregate: A combined measure across multiple cases or groups. / disaggregation: Breaking a combined result into meaningful groups. / severity: The seriousness of a failure's consequences. / false positive: A positive decision when the reference outcome is negative. / false negative: A negative decision when the reference outcome is positive. / confusion matrix: A table comparing predicted and reference categories. / calibration: Agreement between stated confidence and observed outcomes. / confidence interval: A range expressing uncertainty under stated statistical assumptions. / statistical significance: Evidence against a null model under a specified statistical test. / practical significance: The importance of a difference for the actual task. / paired comparison: Comparison of results on corresponding cases. / LLM-as-judge: Use of a language model to evaluate outputs. / human review: Assessment by people using defined criteria. / release gate: A defined decision checkpoint before release.
Speaking task: Check the dialogue. Read both slices: 12 to 18 of 20 long cases; 78 to 70 of 80 short cases. Switch roles. State that total passes fall from 90 to 88; severity and production mixture remain unmeasured. Complete the urgent-message exchange. Name the four misses as false negatives and retain the required severity review.
Partner: Noor: Evaluation manager / Ben: Machine-learning engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier has a high overall score but misses four of ten urgent messages. Urgent-message consequences have not been reviewed. The team must not treat aggregate success as sufficient evidence for release.
Model for comparison AFTER my attempt, not a response to give me first: Ben: I will revise the slide to show both counts, the shared rubric, and the missing traffic and severity analysis. The candidate's strengths will still be visible.
Noor: Good. That is disaggregation, not an attempt to hide the average. We want the decision-maker to see why the headline alone is insufficient.
Ben: I'll recommend waiting for the changed-case review. I still think routing is worth testing, but I won't present it as a fix today.
Noor: Agreed. The release gate remains open. The next update should say which criteria are met, which are not, and what evidence is still missing.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 06: Evaluation, Benchmarks, and Regression
Original fictional case: On the same 100-case test, the previous model passes 90 cases and the candidate passes 88. Among 20 long-document cases, passes rise from 12 to 18. Among 80 short-request cases, passes fall from 78 to 70. The scoring rubric is unchanged. Production traffic proportions and the seriousness of individual failures have not been analyzed. Noor chairs the review; Ben prepared the results.
Language workshop: The average is down, but one slice improved
Communication goal: Present disaggregated results and challenge an oversimplified release argument.
Language guidance: The aggregate falls by two percentage points. Long-document performance rises from 60 to 90 percent; short-request performance falls from 97.5 to 87.5 percent. These are test-set rates, not verified production estimates.
Useful expressions (preserve the supplied facts): Overall passes fell, while the long-document slice improved. / Eighteen of 20 long-document cases passed. / Short-request passes fell from 78 to 70 out of 80. / Which previously passing cases now fail? / A count alone does not tell us how serious those failures are. / Does this test mixture resemble the traffic we plan to serve? / We cannot infer production performance from these counts alone. / Compare the changed cases and classify their failure severity. / The improved slice should not replace the overall result on the slide. / Were both versions scored against the same rubric? / We have not established whether this difference is stable across samples. / Which acceptance criterion applies to the short-request regression? / Let us agree on the criteria before selecting the winning score. / How closely does the automated judge match the reviewed labels? / Routing long documents separately is a proposal to test, not an approved fix. / We will defer the decision until the changed cases have been assessed.
Vocabulary: evaluation: A structured assessment against specified criteria. / benchmark: A reference test used for comparison. / baseline: The reference system or result for a comparison. / candidate: The version under consideration. / regression: A deterioration in previously measured behavior. / test case: A defined input and expected criterion or outcome. / golden set: Curated cases used to check important behavior repeatedly. / pass rate: The fraction of evaluated cases meeting a criterion. / acceptance criterion: A stated condition required for acceptance. / slice: A defined subgroup within evaluation data. / aggregate: A combined measure across multiple cases or groups. / disaggregation: Breaking a combined result into meaningful groups. / severity: The seriousness of a failure's consequences. / false positive: A positive decision when the reference outcome is negative. / false negative: A negative decision when the reference outcome is positive. / confusion matrix: A table comparing predicted and reference categories. / calibration: Agreement between stated confidence and observed outcomes. / confidence interval: A range expressing uncertainty under stated statistical assumptions. / statistical significance: Evidence against a null model under a specified statistical test. / practical significance: The importance of a difference for the actual task. / paired comparison: Comparison of results on corresponding cases. / LLM-as-judge: Use of a language model to evaluate outputs. / human review: Assessment by people using defined criteria. / release gate: A defined decision checkpoint before release.
Speaking task: Check the dialogue. Read both slices: 12 to 18 of 20 long cases; 78 to 70 of 80 short cases. Switch roles. State that total passes fall from 90 to 88; severity and production mixture remain unmeasured. Complete the urgent-message exchange. Name the four misses as false negatives and retain the required severity review.
Partner: Noor: Evaluation manager / Ben: Machine-learning engineer
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A classifier has a high overall score but misses four of ten urgent messages. Urgent-message consequences have not been reviewed. The team must not treat aggregate success as sufficient evidence for release.
Model for comparison AFTER my attempt, not a response to give me first: Ben: I will revise the slide to show both counts, the shared rubric, and the missing traffic and severity analysis. The candidate's strengths will still be visible.
Noor: Good. That is disaggregation, not an attempt to hide the average. We want the decision-maker to see why the headline alone is insufficient.
Ben: I'll recommend waiting for the changed-case review. I still think routing is worth testing, but I won't present it as a fix today.
Noor: Agreed. The release gate remains open. The next update should say which criteria are met, which are not, and what evidence is still missing.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Retrieve occupational vocabulary
1. Select four terms from the reference. Explain each in plain English, give two natural collocations and quote or create a short workplace example. Label invented examples. Contrast each term with a plausible confusable term; acknowledge regional variation.
2. Present one role-labelled exchange with one missing word or phrase and four choices A-D. Exactly one choice must fit the stated facts and grammar. Stop and wait for my answer before revealing it.
3. Explain all four options, then invite me to read the completed exchange aloud. Offer a sentence frame for optional spoken practice, never an open-ended writing assignment.
4. Repeat for the other three terms, one question at a time. Vary the correct letter and the situation; preserve the reference facts. If a choice is ambiguous, replace the item.
5. Finish with two new four-choice collocation questions for later review, with a clearly separate answer key. Do not claim to schedule a reminder.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 07: Inference, Latency, Cost, and Deployment
Original fictional case: A smaller model reduces mean response time from four seconds to two on a fixed test. Complex-answer quality and peak-load tail latency remain unchecked. Maya owns serving operations; Luis owns the product pilot. An internal pilot may use synthetic requests only. A production rollout requires a separate review. The previous configuration remains available, but rollback ownership has not been assigned.
Language workshop: Faster on average is not ready for everyone
Communication goal: Negotiate a staged deployment with explicit conditions and rollback ownership.
Language guidance: A mean of two seconds does not mean every request finishes within two seconds. Time to first token and time to complete answer measure different experiences. Streaming can change one without reducing the other.
Useful expressions (preserve the supplied facts): Mean response time fell from four seconds to two on this test. / Complex-answer quality is still unmeasured. / What happens to the slowest requests under peak load? / Is that time to first token or time to complete the answer? / The internal pilot uses synthetic requests only. / We need an agreed condition for pausing the pilot. / Who can initiate rollback if the condition is met? / Pilot approval does not authorize production rollout. / Were both models tested under the same demand and input mix? / Please include retries and fallback traffic in the cost comparison. / How much of the delay occurs before inference begins? / I can provide the load-test results, not promise a two-second maximum. / Has the previous configuration been checked for restoration? / Record the owner, contact route, and approved recovery procedure. / Show the distribution as well as the mean. / We can proceed with the bounded test once the owners and conditions are confirmed.
Vocabulary: latency: Time taken for a request or processing step. / mean: The arithmetic average of observed values. / median: The middle value in an ordered set. / P95: The 95th percentile of a measured distribution. / tail latency: Response times toward the slow end of a distribution. / throughput: The rate at which work is completed. / concurrency: Work occurring at the same time. / queueing: Waiting for resources before processing. / batching: Processing multiple requests together. / streaming: Sending portions of output before the full response is complete. / time to first token: Delay before the first generated token arrives. / token rate: The rate of generated or processed tokens. / autoscaling: Automatic adjustment of resource capacity. / cold start: Initial startup delay before a resource is ready. / load test: A test under specified levels of demand. / canary: A limited release used to observe behavior before expansion. / rollout: Gradual or full introduction of a configuration. / rollback: Return to a previous configuration. / fallback: An alternative behavior when the preferred path cannot serve. / SLO: Service-level objective; a defined internal service target. / error budget: The permitted unreliability associated with a service objective. / rate limit: A restriction on request frequency or volume. / capacity planning: Estimating resources needed for expected demand. / cost per request: Allocated processing expense for a defined request unit.
Speaking task: Check the answers. Read four seconds to two as a 50-percent reduction in the measured mean, not a maximum for every request. Switch roles. Keep the pilot limited to synthetic requests and ask who owns rollback; availability of the old configuration is not an assigned owner. Complete the timing exchange. Say one second to first token and six seconds to completion, with neither treated as an approved service target.
Partner: Maya: Serving engineer / Luis: Product lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A demonstration shows the first token after one second, but the complete answer takes six seconds. A slide calls the response time one second. No service target has been approved.
Model for comparison AFTER my attempt, not a response to give me first: Luis: Can you provide the distribution, complex-answer results, and cost assumptions by Friday? I can postpone the external claim until we have that package.
Maya: I can provide the test results and unresolved items. If queueing dominates at higher demand, I will say that rather than attribute every delay to the model.
Luis: Good. The pilot note will state the synthetic-only population, the pause owner, the proposed restoration owner, and the review required before we begin.
Maya: Then we have a bounded plan. We will report the measured gain alongside its limitations, and let the evidence determine the next step instead of promising a universal two-second response.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Choose grammar that preserves the meaning
1. Use the lesson's language workshop, or one actual sentence from the dialogue, to explain a single grammar distinction in no more than 90 words. Show how it affects timing, certainty, responsibility or politeness.
2. Give one brief occupational exchange containing a gap and four complete options A-D. Supply enough context for exactly one answer. Keep distractors plausible; distinguish grammar errors from wording that changes the facts. Stop and wait.
3. Explain why each choice fits or fails, then ask me to read the completed exchange aloud. Do not request a draft, original written sentence or essay.
4. Repeat with three new exchanges, one at a time, moving from recognition to a more subtle meaning contrast. Do not copy the published checks. Label invented variations and do not turn uncertainty into a promise.
5. End with a short contrast table and two supported spoken turns using supplied frames. Accept accurate alternatives in oral practice; acknowledge and replace ambiguous multiple-choice items.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 07: Inference, Latency, Cost, and Deployment
Original fictional case: A smaller model reduces mean response time from four seconds to two on a fixed test. Complex-answer quality and peak-load tail latency remain unchecked. Maya owns serving operations; Luis owns the product pilot. An internal pilot may use synthetic requests only. A production rollout requires a separate review. The previous configuration remains available, but rollback ownership has not been assigned.
Language workshop: Faster on average is not ready for everyone
Communication goal: Negotiate a staged deployment with explicit conditions and rollback ownership.
Language guidance: A mean of two seconds does not mean every request finishes within two seconds. Time to first token and time to complete answer measure different experiences. Streaming can change one without reducing the other.
Useful expressions (preserve the supplied facts): Mean response time fell from four seconds to two on this test. / Complex-answer quality is still unmeasured. / What happens to the slowest requests under peak load? / Is that time to first token or time to complete the answer? / The internal pilot uses synthetic requests only. / We need an agreed condition for pausing the pilot. / Who can initiate rollback if the condition is met? / Pilot approval does not authorize production rollout. / Were both models tested under the same demand and input mix? / Please include retries and fallback traffic in the cost comparison. / How much of the delay occurs before inference begins? / I can provide the load-test results, not promise a two-second maximum. / Has the previous configuration been checked for restoration? / Record the owner, contact route, and approved recovery procedure. / Show the distribution as well as the mean. / We can proceed with the bounded test once the owners and conditions are confirmed.
Vocabulary: latency: Time taken for a request or processing step. / mean: The arithmetic average of observed values. / median: The middle value in an ordered set. / P95: The 95th percentile of a measured distribution. / tail latency: Response times toward the slow end of a distribution. / throughput: The rate at which work is completed. / concurrency: Work occurring at the same time. / queueing: Waiting for resources before processing. / batching: Processing multiple requests together. / streaming: Sending portions of output before the full response is complete. / time to first token: Delay before the first generated token arrives. / token rate: The rate of generated or processed tokens. / autoscaling: Automatic adjustment of resource capacity. / cold start: Initial startup delay before a resource is ready. / load test: A test under specified levels of demand. / canary: A limited release used to observe behavior before expansion. / rollout: Gradual or full introduction of a configuration. / rollback: Return to a previous configuration. / fallback: An alternative behavior when the preferred path cannot serve. / SLO: Service-level objective; a defined internal service target. / error budget: The permitted unreliability associated with a service objective. / rate limit: A restriction on request frequency or volume. / capacity planning: Estimating resources needed for expected demand. / cost per request: Allocated processing expense for a defined request unit.
Speaking task: Check the answers. Read four seconds to two as a 50-percent reduction in the measured mean, not a maximum for every request. Switch roles. Keep the pilot limited to synthetic requests and ask who owns rollback; availability of the old configuration is not an assigned owner. Complete the timing exchange. Say one second to first token and six seconds to completion, with neither treated as an approved service target.
Partner: Maya: Serving engineer / Luis: Product lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A demonstration shows the first token after one second, but the complete answer takes six seconds. A slide calls the response time one second. No service target has been approved.
Model for comparison AFTER my attempt, not a response to give me first: Luis: Can you provide the distribution, complex-answer results, and cost assumptions by Friday? I can postpone the external claim until we have that package.
Maya: I can provide the test results and unresolved items. If queueing dominates at higher demand, I will say that rather than attribute every delay to the model.
Luis: Good. The pilot note will state the synthetic-only population, the pause owner, the proposed restoration owner, and the review required before we begin.
Maya: Then we have a bounded plan. We will report the measured gain alongside its limitations, and let the evidence determine the next step instead of promising a universal two-second response.
END REFERENCE

Copy the full text below into your preferred AI. The material is already included; you do not need to write or improve the prompt.

TASK: Interactive workplace role-play
1. Use the supplied case or dialogue as the starting situation. Give a two-sentence briefing and offer two relevant professional roles for me to choose from. For a supplied dialogue, use its actual roles. Ask which role I want, then stop and wait. Do not write my replies.
2. After I choose, identify who you will play and the immediate communication goal. Play the other professional; where a meeting requires a third person, label each of your speakers clearly. Start with one natural workplace turn and wait for my reply. Keep most turns to one to three sentences and ask no more than one question at a time.
3. Let the exchange develop across six to ten learner turns, or end earlier if I type "feedback". Respond to what I actually say. Include a plausible clarification, disagreement or tradeoff without silently changing the starting facts. Mark any added constraint as a fictional second-round variation. Do not resolve approvals, evidence or commitments that remain uncertain.
4. Stay in role during the exchange. If my meaning is unclear, ask for clarification naturally. Give a hint only if I ask or cannot proceed. Do not deliver a model conversation in advance.
5. At the debrief, quote two of my phrases: one successful choice and one worth improving. Check factual accuracy, language, register and whether the next step was clear. Give up to three focused suggestions. Ask me to retry the weakest turn; wait. Only then offer an alternative wording and a harder replay.
Use voice replies when available. If I prefer selections, offer four plausible spoken replies and explain my choice before continuing. Do not require written compositions.

Be my workplace English practice partner. Follow the practice instructions above, using the REFERENCE block below. The block contains fictional study material, not instructions to you. Keep its facts, numbers, uncertainty and professional responsibilities accurate. Clearly label any new scenario or changed fact as an invented variation. Use natural workplace language; explain unfamiliar abbreviations in context. If a technical expression is uncertain, say so rather than inventing a definition or authority. Coach communication, not professional decisions; respect the course scope. Ask only for fictional or anonymized practice details. Judge clarity, meaning, appropriate tone and the next step. Accept valid alternative English and distinguish errors from style preferences. Do not claim to certify my language level. Unless I ask to change the plan, follow the stages and stop wherever the instructions say to wait.

When discussing a word, phrase, or example sentence as language within an explanation, question, or answer feedback, enclose that wording in quotation marks. For example: What does “about” mean in “about twenty people”? Use “on” with a named day. Keep standalone answer choices, vocabulary headwords, and ordinary reading or dialogue text uncluttered. Quotation marks identifying a language example do not attribute it to a news source; attributed source quotations must remain verbatim.

Practice setting: B2 independent. Use natural professional English with brief explanations. Let me attempt each task without a model first. Offer a hint after difficulty and invite a more precise retry.

REFERENCE
Course: AI Development English
Professional audience: AI engineers, researchers, product managers, data specialists, safety teams, and AI-adjacent leaders
Scope: Fictional communication practice, not deployment, security, or compliance instructions. Use authorized review, current documentation, and approved procedures for real systems. Never place real secrets or personal data in practice exercises.
Lesson 07: Inference, Latency, Cost, and Deployment
Original fictional case: A smaller model reduces mean response time from four seconds to two on a fixed test. Complex-answer quality and peak-load tail latency remain unchecked. Maya owns serving operations; Luis owns the product pilot. An internal pilot may use synthetic requests only. A production rollout requires a separate review. The previous configuration remains available, but rollback ownership has not been assigned.
Language workshop: Faster on average is not ready for everyone
Communication goal: Negotiate a staged deployment with explicit conditions and rollback ownership.
Language guidance: A mean of two seconds does not mean every request finishes within two seconds. Time to first token and time to complete answer measure different experiences. Streaming can change one without reducing the other.
Useful expressions (preserve the supplied facts): Mean response time fell from four seconds to two on this test. / Complex-answer quality is still unmeasured. / What happens to the slowest requests under peak load? / Is that time to first token or time to complete the answer? / The internal pilot uses synthetic requests only. / We need an agreed condition for pausing the pilot. / Who can initiate rollback if the condition is met? / Pilot approval does not authorize production rollout. / Were both models tested under the same demand and input mix? / Please include retries and fallback traffic in the cost comparison. / How much of the delay occurs before inference begins? / I can provide the load-test results, not promise a two-second maximum. / Has the previous configuration been checked for restoration? / Record the owner, contact route, and approved recovery procedure. / Show the distribution as well as the mean. / We can proceed with the bounded test once the owners and conditions are confirmed.
Vocabulary: latency: Time taken for a request or processing step. / mean: The arithmetic average of observed values. / median: The middle value in an ordered set. / P95: The 95th percentile of a measured distribution. / tail latency: Response times toward the slow end of a distribution. / throughput: The rate at which work is completed. / concurrency: Work occurring at the same time. / queueing: Waiting for resources before processing. / batching: Processing multiple requests together. / streaming: Sending portions of output before the full response is complete. / time to first token: Delay before the first generated token arrives. / token rate: The rate of generated or processed tokens. / autoscaling: Automatic adjustment of resource capacity. / cold start: Initial startup delay before a resource is ready. / load test: A test under specified levels of demand. / canary: A limited release used to observe behavior before expansion. / rollout: Gradual or full introduction of a configuration. / rollback: Return to a previous configuration. / fallback: An alternative behavior when the preferred path cannot serve. / SLO: Service-level objective; a defined internal service target. / error budget: The permitted unreliability associated with a service objective. / rate limit: A restriction on request frequency or volume. / capacity planning: Estimating resources needed for expected demand. / cost per request: Allocated processing expense for a defined request unit.
Speaking task: Check the answers. Read four seconds to two as a 50-percent reduction in the measured mean, not a maximum for every request. Switch roles. Keep the pilot limited to synthetic requests and ask who owns rollback; availability of the old configuration is not an assigned owner. Complete the timing exchange. Say one second to first token and six seconds to completion, with neither treated as an approved service target.
Partner: Maya: Serving engineer / Luis: Product lead
Guided response task: Select supplied answers for the dialogue gaps. Check the explanations, then read the completed exchange aloud; no original writing is required.
Optional second-round challenge: A demonstration shows the first token after one second, but the complete answer takes six seconds. A slide calls the response time one second. No service target has been approved.
Model for comparison AFTER my attempt, not a response to give me first: Luis: Can you provide the distribution, complex-answer results, and cost assumptions by Friday? I can postpone the external claim until we have that package.
Maya: I can provide the test results and unresolved items. If queueing dominates at higher demand, I will say that rather than attribute every delay to the model.
Luis: Good. The pilot note will state the synthetic-only population, the pause owner, the proposed restoration owner, and the review required before we begin.
Maya: Then we have a bounded plan. We will report the measured gain alongside its limitations, and let the evidence determine the next step instead of promising a universal two-second response.
END REFERENCE