apptek-com/recall-rewrite-oasst1

Recall Rewrite OASST1: knowledge-aligned SFT data

Data release for the paper "Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning" (Becker, Kemmler, Thulke, Schäfer, Dugast, Ney; accepted at EMNLP 2026, Main Conference).

Knowledge-aligned SFT constrains supervised fine-tuning targets to what the base model already knows. Recall Rewrite implements this without external evidence: every gold response of the SFT set is decomposed into atomic claims, each knowledge-dependent claim is turned into five probing questions, the base model answers each question twice, and a claim is kept only if the base model consistently recalls it. Claims it cannot recall are removed by a rewriter, or the whole response is replaced by a refusal if too little remains.

This repository contains

• the resulting knowledge-aligned training sets for Qwen/Qwen3-4B-Base and allenai/Olmo-3-1025-7B (*-sft configs, ready for training), and
• all intermediate pipeline outputs (*-full configs): decomposed claims, the knowledge-dependence audit, the probing questions, the base-model answers with their entailment labels, and the resulting known/unknown labels. This allows to re-derive the training set under other filter thresholds offline.

Quick start

from datasets import load_dataset

# flat SFT data: prompt / response / messages
sft = load_dataset("apptek-com/recall-rewrite-oasst1", "qwen3-4b-base-sft", split="train")

# everything the pipeline produced
full = load_dataset("apptek-com/recall-rewrite-oasst1", "qwen3-4b-base-full", split="train")
ex = full[0]
claim = next(c for c in ex["claims"] if c["is_knowledge_dependent"])
print(claim["text"], claim["is_known"], claim["probes"][0]["question"], claim["probes"][0]["answers"][0]["label"])

Files are plain JSONL (gzip-compressed where large), so jq or pandas.read_json(lines=True) work as well.

Source data and pipeline

• SFT data. OASST1 (Köpf et al., 2023), English, first turn of each conversation tree, highest-ranked response: 3,468 prompt–response pairs. One example is dropped by the pipeline because its claim decomposition could not be parsed, leaving 3,467 examples (identical across the two base models).
• Claim decomposition and probing questions were generated once with gpt-5-mini-2025-08-07 and are shared by both base models: 64,301 claims, of which 39,853 are knowledge-dependent and carry 5 probing questions each. The remaining 24,448 claims (opinions, instructions, context-dependent statements, ...) are always kept.
• Probe answers: 2 samples per question from the base model (temperature 0.5, few-shot QA prompt), labelled ENTAILS / CONTRADICTS / UNRELATED against the claim by gpt-5-mini-2025-08-07.
• Decision rule (paper Sec. 3.2.2, thresholds $j_e/k_e/j_c/k_c$): a knowledge-dependent claim is known iff at least $j_e$ questions have at least $k_e$ entailing answers and at most $j_c$ questions have at least $k_c$ contradicting answers. The released rewrites use the paper default 2/1/2/1.
• Rewriting: gpt-5-mini-2025-08-07 removes the unknown claims with minimal edits. If too little information remains, the response becomes the fixed refusal "I cannot answer as I have insufficient information." (is_refusal = true).

The paper's Recall Rewrite models are trained on all records of the corresponding *-sft config (3,165 non-refusals + 302 refusals for Qwen; 3,175 + 292 for OLMo) with the messages field and the base model's chat template, no system prompt. The ablations in the paper subsample refusals or filter differently; those variants are not shipped as separate files but can be recreated from the *-full configs (see known_claim_rule.py in the GitHub repository).

Statistics (paper Table 7)

	Qwen3-4B-Base	OLMo-3-7B
Examples	3,467	3,467
Claims / knowledge-dependent	64,301 / 39,853	64,301 / 39,853
Probe answers ENTAILS / UNRELATED / CONTRADICTS	62.5% / 25.1% / 12.4%	70.6% / 15.3% / 14.1%
Known claims (kept)	31,654 (79.4%)	32,168 (80.7%)
Unknown claims (removed)	8,199 (20.6%)	7,685 (19.3%)
Non-refusal / refusal responses	3,165 / 302	3,175 / 292

Configs and schema

<model>-sft (qwen3-4b-base-sft, olmo-3-7b-sft)

One record per OASST1 example, 3,467 rows.

field	type	description
id	string	oasst1-NNNN, identical across the two base models and the full configs
base_model	string	HF id of the base model whose knowledge the target was aligned to
prompt	string	OASST1 user prompt
response	string	knowledge-aligned target (rewritten OASST1 response, or the refusal string)
is_refusal	bool	target is the refusal string
messages	list of {role, content}	the same as a two-turn chat, as used for training

<model>-full (qwen3-4b-base-full, olmo-3-7b-full)

Same rows, with everything the pipeline produced.

field	type	description
id, base_model, prompt, is_refusal		as above
original_response	string	OASST1 gold response
rewritten_response	string	knowledge-aligned target (= response in the sft config)
filter	string	thresholds $j_e/k_e/j_c/k_c$ used for the labels and the rewrite ("2/1/2/1")
num_claims, num_knowledge_dependent_claims, num_known_claims, num_unknown_claims	int	per-response counts
claims[]	list of struct	one entry per atomic claim of original_response
claims[].claim_id	int	position in the decomposition
claims[].text	string	the atomic claim
claims[].is_knowledge_dependent	bool	passed the audit (the paper's knowledge-dependent claim); only these are probed
claims[].knowledge_dependence_reason	string	the auditor's final justification
claims[].audit_checklist	struct of 4 strings	is_independent_of_context, contains_specific_fact, is_not_obvious, can_be_asked_for
claims[].is_known	bool or null	decision-rule result; null for non-knowledge-dependent claims
claims[].is_removed	bool	claim was handed to the rewriter for removal (= knowledge-dependent and not known)
claims[].num_entailed_answers, num_contradicted_answers, num_unrelated_answers	int	label counts over the 10 probe answers (0 for non-probed claims)
claims[].probes[]	list of struct	5 entries for knowledge-dependent claims, empty otherwise
claims[].probes[].question	string	probing question
claims[].probes[].answers[]	list of {text, label}	2 base-model answers with label ENTAILS / CONTRADICTS / UNRELATED

Note that claims, probes[].question, is_knowledge_dependent and the audit fields are identical between the two base models; only the answers, labels, is_known/is_removed and the rewrites differ.

Caveats

• Everything except the OASST1 prompts and responses is model-generated (gpt-5-mini for claims, questions, labels and rewrites; the base models for the probe answers). Labels are noisy; the paper's qualitative appendix discusses typical failure modes (over-truncation, incoherent rewrites, refusals of code-heavy responses).
• The known/unknown labels describe what one specific base model recalled under one probing setup (5 questions × 2 samples, temperature 0.5). They are not ground-truth facts about the world and not transferable to other models.

License

The dataset is released under CC BY-SA 4.0. The prompts and original responses come from OASST1 (Köpf et al., 2023, Apache-2.0). The rewritten responses, claims, probing questions and labels were generated with gpt-5-mini and the respective base models.

Citation

@inproceedings{becker2026stick,
  title     = {Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning},
  author    = {Becker, Arthur and Kemmler, Jakob and Thulke, David and Sch{\"a}fer, Christine and Dugast, Christian and Ney, Hermann},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  month     = oct,
  year      = {2026},
  address = "Budapest, Hungary"
}

Contact

AppTek

• dthulke@apptek.com