ChenWu98/grafting-data

Grafting data

Data for the paper Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning. Code: ar-forum/grafting.

Math

Each model has a train split of off-policy SFT data on hard problems the post-trained model cannot solve, and a disjoint heldout split of hard problems used to measure new-task performance.

config	train	heldout	training data
math_qwen3_1p7b_thinking	1,533	111	teacher traces (AoPS subset of Nemotron-SFT-Math), problems with 0/16 solves
math_qwen3_1p7b_thinking_star	1,093	(same as above)	STaR: one verified self-generated trace per solved problem
math_qwen3_8b_thinking	985	100	teacher traces from the same pool, problems with at most 1/16 solves
math_qwen3_8b_nonthinking	1,163	200	worked solutions to DAPO problems Qwen3-8B (non-thinking) never solves
math_olmo3_7b_thinking	730	100	teacher traces from the same pool, problems OLMo-3-7B (thinking) never solves in 16 attempts

Columns: problem (raw problem), question (the user message we use, i.e. problem plus "Please reason step by step, and put your final answer within \boxed{}."), response (SFT target, train only), answer (reference answer).

Old-task performance in the paper is the mean over AIME 2024, AIME 2025 and HMMT 2025 (30 problems each), in the old_task config with the same columns as heldout. We sample 16 generations per problem at temperature 1.0 and top-p 1.0 (up to 32K tokens in thinking mode) and report the mean score, using the scorer in the code repository.

Knowledge

New facts about 200 fictional entities (1,125 atomic facts), so the facts are absent from any pretrained model.

config / split	rows	what
knowledge / train	9,000	QA pairs, 8 phrasings per fact
knowledge / train_mix	10,533	the 9,000 QA pairs plus the 1,533 math traces of math_qwen3_1p7b_thinking; all methods train on this mix
knowledge / heldout	2,250	held-out paraphrased questions, never seen in training
knowledge_articles	200	wiki-style articles stating the facts, used only to condition the OPSD teacher

Answers are given inside <answer></answer> tags. We decode greedily and score by exact match of the text inside the tags against answer.

Retention rollouts

retention_qwen3_1p7b has 1,858 responses sampled from Qwen3-1.7B (thinking) at temperature 1.0, one per problem. It is the --data input for the Fisher estimate (python -m grafting.fisher) used to build the Fisher mask.

Citation

@article{wu2026grafting,
  title   = {Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning},
  author  = {Wu, Chen Henry and Zhang, Thomas and Raghunathan, Aditi},
  journal = {arXiv preprint arXiv:2610.05872},
  year    = {2026},
}