IncorrectTraceSFT · MATH random mask 90

A full fine-tune of Qwen2.5-7B from How Should Incorrect Traces Be Used in Supervised Fine-Tuning? This checkpoint is the random_mask_90 condition in the math masking experiment.

Code and reproduction instructions · Training data

Training

The student starts from Qwen/Qwen2.5-7B, revision d149729398750b98c0af14eb82c78cfe92750796. Training uses SLIME with Megatron-LM, BF16, global batch size 64, and AdamW with a peak learning rate of 5e-6. Full settings and exact data selections are in the code repository.

This is the final checkpoint after 125 optimizer updates. The original checkpoint identifier is 37ef3bef13981e172d7a5df1325be06d9e283039f0e5e829094719998e81b023.

Evaluate

From a checkout of the code repository with evaluation dependencies installed:

python -m incorrect_trace_sft.evaluate --experiment math_masking --model suryadv/IncorrectTraceSFT-Qwen2.5-7B-MATH-random-mask-90 --output runs/MATH-random-mask-90

The evaluation command preserves the experiment's prompts, context limit, sampling parameters, and answer scorer. It reports pass@1 and the Monte Carlo standard error across decoding repeats. See the repository's results table for all conditions and comparisons.

Files and intended use

The safetensors weights are unchanged from the evaluated checkpoint. provenance.json records their original SHA-256 hashes. The tokenizer configuration has been normalized for Transformers 4.57: legacy special-token metadata is represented using the current field format, and the existing chat template is embedded. Vocabulary, token IDs, and the chat template are unchanged.

This model is released for research on mathematical reasoning and SFT data selection. Correct final answers in the training data do not imply that intermediate reasoning steps were verified. ID/OOD in this study refers to problem identity relative to SFT; it says nothing about pretraining exposure. Coding-agent checkpoints will be released separately.