fine-tuning-base-model-instruction-data-removing-training-targets-any CLM-0048When fine-tuning a base model on instruction data, removing from the training targets any factual claim the base model cannot consistently recall raises the supported-claim rate of later long-form generations, but the gain comes largely from a more conservative response policy — more refusals and fewer supported claims per answer — not from generating more correct facts at equal coverage.
Capability: Stating false facts confidently
Observed on
Shown on Qwen3-4B-Base and OLMo 3 7B fine-tuned on the small English first-turn OASST1 set, evaluated on entity-centric long-form factuality benchmarks (WildHalu, Biography) with a.
Sources
- Compared four knowledge-alignment methods against standard SFT under one training setup, plus an ablation varying only the share of known claims (100/50/0%) with prompt count and refusal count fixed, which is the part that isolates causality. Factuality metrics count refusals as fully supported, so the headline gain is partly definitional; the authors say so and report the falling supported-claim counts. Single dataset, small SFT corpus, automatic verifiers.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When injecting new or counterfactual facts into a 3-4B instruct model by post-training, supervised fine-tuning on extracted QA pairs binds facts to their training question form — recall of the single fact stays reasonable while multi-source and inferential questions over the same facts collapse — and adding more SFT epochs does not close the gap, whereas mixed-policy RL that supplies the golden answer as an off-policy trajectory when on-policy rollouts fail recovers much of the generalization.Absorbing and flexibly using new knowledge · reviewed by AI
- When training small instruction-tuned models with reinforcement learning plus factual supervision, routing each atomic fact's verification score only to the tokens that produced it, and down-weighting verifier judgements that do not change when their key evidence is removed, improves factuality benchmark scores over trajectory-level or reasoning- step-level factual rewards.Checking claims against evidence · reviewed by AI
- In LLM fact-checking over supplied evidence, high veracity accuracy does not imply the verdict depends on that evidence: when the model's own cited sentences are iteratively removed, off-the-shelf models (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep their SUPPORT/REFUTE labels at a non-zero plateau instead of abstaining, and training with paired supervision that requires NOT_ENOUGH_INFO on the evidence-ablated input removes most of that persistence without losing accuracy.Checking claims against evidence · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A run where recall-filtered targets raise supported-claim percentage while holding refusal count and number of supported claims per response equal to standard SFT would show the gain is not coverage-driven; conversely, showing no factuality gain at all when the share of known claims in targets is varied would falsify the mechanism. Proposed technique, not catalogued: Recall-consistency filtering of fine-tuning targets.