Datasetpublic

Ram20307/slm-reasoning-dpo-pairs

SLM Reasoning Research — DPO preference pairs Part of the SLM Reasoning Research project. Preference pairs for DPO training: chosen = Arm A's teacher trace (verified correct), rejected = the untrained Qwen3-0.6B-Base model's own natural wrong attempt on that same GSM8K train question — not a synthetic negative. 7,435 train questions attempted; pairs kept only for the questions the base model got wrong (39.4%). Each row: example_id, prompt, chosen, rejected, reference_answer. See… See the full description on the dataset page: https://huggingface.co/datasets/Ram20307/slm-reasoning-dpo-pairs.

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face

0likes40downloads

Dataset Card

SLM Reasoning Research — DPO preference pairs

Part of the SLM Reasoning Research project.

Preference pairs for DPO training: chosen = Arm A's teacher trace (verified correct), rejected = the untrained Qwen3-0.6B-Base model's own natural wrong attempt on that same GSM8K train question — not a synthetic negative. ~7,435 train questions attempted; pairs kept only for the questions the base model got wrong (~39.4%).

Each row: example_id, prompt, chosen, rejected, reference_answer.

See the project repo for how raw-DPO and SFT+DPO checkpoints trained on this data compare.