Design a fine-tune that teaches a model to refuse when the retrieved context lacks the answer.

Short answer·Medium·4.0 · 0·~3 min·Relevant atAnthropic Google OpenAI

Attempt it

A RAG product is shipping confident hallucinations whenever the retriever misses. Design an SFT recipe that teaches the model to refuse, i.e. say 'I don't have enough information in the provided context', when the retrieved chunks do not actually support an answer. Specify the data shape, how you construct positive vs negative examples, the balance between them, and at least one failure mode to watch.

Free · 2 AI evals / day

**The data shape** is triples of (question, retrieved_chunks, gold_response). Build the dataset in two halves of roughly equal size. **The positive half** contains examples where the chunks genuinely support an answer. The gold response uses the chunks to answer, ideally citing the supporting passage. This teaches grounded behaviour. **The negative half** contains examples where the chunks are plausible but do not actually answer the question. The gold response is a templated refusal like 'I don't have enough information in the provided context'. This teaches the model to refuse on insufficient context. The critical detail is that the negatives must look realistic. If you feed empty context or random off-topic text as negatives, the model learns to refuse only on those obvious cases. Hard production cases keep producing hallucinations. Source realistic negatives by perturbing the query slightly before retrieval, swapping in chunks from a neighbouring question, or using the production retriever's actual misses. Keep the refusal phrasing templated and small (one or two wordings) so the model learns the cue, not the phrasing. Add 10 to 15% partially-supported cases with hedged responses to avoid all or nothing behaviour. Evaluate refusal precision and recall separately on a held-out set with supported, unsupported, and partial cases.

**The recipe** is (question, retrieved_chunks, gold_response) triples split into two matched populations. **Positive half**: chunks support the answer, gold response uses them. Source from real production successes where the chunks contained the answer. **Negative half**: chunks do not support the answer, gold response is a templated refusal. Construction of these negatives is the make or break decision. **Sourcing realistic negatives.** Three techniques. Perturb the query slightly before retrieval (paraphrase, swap entity) so the retriever returns plausibly-related but incorrect chunks. Swap a top-k slot from a different question. Mine the production retriever's actual misses where users gave thumbs-down. If you use empty or wildly off-topic chunks, the model only learns to refuse on those obvious shapes. The subtle near-miss cases that drive production hallucinations still hallucinate. **Refusal templating.** Use one or two canonical phrasings consistently. The signal the model should learn is 'context does not support, refuse', not 'memorise these words'. **Partial-support slice.** Mix in 10 to 15% of cases where chunks address part of the question. Gold response hedges. Without this slice the model collapses to binary answer or refuse. **Evaluation.** Track refusal precision and recall separately on a held-out set with supported, unsupported, and partial cases. Overall accuracy hides the trade-off. **Failure modes**: over-refusal (negatives too aggressive), under-refusal (negatives too easy), and phrasing drift (templating too loose).

Quick reference

TL;DR

Build matched (question, chunks, response) triples: half where chunks support the answer, half where chunks look related but do not, with templated refusals. Realistic negatives matter most.

Memory aid

Easy to grasp

Picture teaching a kid to say 'I don't know' when their notes do not cover a question. You quiz them with two kinds of cards. On half the cards, the notes do answer the question, and the kid is rewarded for using them. On the other half, the notes look related but actually do not answer it, and the kid is rewarded for saying 'these notes do not cover that'. The trick is the second pile has to look real. If you only quiz them with blank notes, they only learn to refuse when the page is empty, and they keep making things up whenever the page has any text on it. The realistic near-miss notes are what teach the actual signal.

Key concepts

Hallucination Retrieval-augmented generation Grounded refusal

Detailed answer & concept explanation

~10 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

3 min: describe the triple shape, the two matched halves plus partial slice, the three negative-sourcing techniques, the templating discipline, and the precision-recall evaluation.

Real-world examples

Real products, models, and research that use this idea.

Anthropic's Claude Opus 4.7 training reportedly includes grounded-refusal data for RAG and tool-use scenarios, with templated refusal phrasings.
OpenAI's GPT-5.5 RAG fine-tuning guide recommends a similar matched-population recipe, with explicit examples of perturbed-query negatives.
Llama 4 Maverick has been fine-tuned by enterprise teams in 2026 for grounded customer-support assistants using exactly this triple structure on Hugging Face TRL.
Production RAG products at Notion AI and Glean use grounded-refusal fine-tunes to suppress confident hallucinations on retrieval misses, with internal eval sets that track refusal precision and recall.
Allen AI's Tulu 3 data mixture includes a grounded-refusal slice mined from misses against the AI2 reasoning corpus, demonstrating the recipe at open-data scale.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you decide the precision-recall trade-off for refusal at evaluation time?

QWhat would happen if you used the same refusal phrasing in 100 percent of negatives but a varied phrasing in 0 percent of negatives?

QHow would you extend this recipe to handle adversarial users who try to coerce the model into answering despite insufficient context?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Sourcing negative examples as blank or wildly off-topic context. The model then learns to refuse only on those obvious shapes and keeps hallucinating on the subtle near-miss cases that actually hurt in production.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.