Sangsang/rewind-run1-sft-data

Rewind run1 SFT dataset

The final merged SFT training set used for Sangsang/rewind-run1-sft-model. It contains 584 examples from 307 distinct questions: 146 trigger, 146 post-rewind, 146 negative, and 146 clean rows. This is the actual training set, rather than the larger pools of candidate solutions or unfinished recovery shards.

Source preparation sampled four solutions each to 1,000 DAPO-Math problems using Qwen/Qwen3-1.7B in thinking mode, then searched failed traces for sustained drops in estimated continuation success. A rewind deletes four reasoning sections.

Fields

Field	Meaning
question	Original math question
answer	Reference answer used for verification
prefix	Reasoning supplied before the supervised completion
completion	Training target; <rewind> for triggers
type	trigger, post_rewind, negative, or clean
enable_thinking	Whether the Qwen3 thinking template is used
b	Section index where present; nullable

Trigger examples teach the action marker. Post-rewind examples teach a successful continuation from the retained prefix. Negative examples supply successful continuations at on-track positions. Clean examples are complete correct solutions from an empty prefix.

from datasets import load_dataset

data = load_dataset("Sangsang/rewind-run1-sft-data", split="train", token=True)

This repository is private; authenticate with an account that has access. data_summary.json records generation counts, and manifest.json provides file checksums.