midReportedcurated
Tell me about a time a model surprised you in production. How did you find out, and what did you change?
30-second answer
- I treated it as a pattern question first, sampling 200 production traces to confirm the behaviour was systematic before changing anything.
- The traces pointed at two epochs on a narrow single-turn dataset plus a chat template that always carried the JSON schema, so the model learned the envelope as part of every reply.
- I fixed the cause rather than the example, cutting to one epoch and mixing in general instruction data, which cost about two points of task F1 but restored conversational behaviour.
- Verification was on the affected distribution: schema leakage went from 14% to under 1% of off-task turns.
- The process change was to run a general-capability regression suite on every fine-tune and test paraphrased off-distribution prompts before shipping.
- A prompt tweak would have hidden the symptom and left the fine-tune wrapping answers in the schema under any paraphrase.
Likely follow-ups
Open a question to rehearse the answer.
How did you decide it was a pattern rather than an outlier?
- I counted the behaviour across 200 sampled traces and stratified by intent, because a rate by slice separates a systematic template from a one-off turn.
- The clustering on off-task questions was the signal, since a random outlier would not concentrate in the same kind of turn.
- One complaint would have been treated as a reproduction attempt, not evidence, until the sample showed the same shape again.
What test would you write first next time?
- The first test would be a set of off-task paraphrases whose correct output is prose, asserting that no schema wrapper appears.
- I would also add a small general-capability suite before and after the fine-tune, to catch forgetting that task metrics cannot see.
- I would gate on both, because they test different failures, and the gate has to run on the exact chat template used in serving.
Red flags
- Fixes the single reported example with a prompt tweak
- Cannot say how widespread the issue was or how they measured it
Concepts this tests
Tags
Source: codesnatch.io
More in Behavioural & project deep-dives
- Walk me through an AI feature you shipped end to end and your exact role. What decisions were yours?
- Describe a time you disagreed with a product or engineering stakeholder about using an agent, fine-tuning, or RAG.
- Tell me about a model-quality regression you introduced or discovered after deployment. What did you do?