qwen2.5-3b-math-sft-203

Qwen2.5-3B Math SFT 203

Qwen2.5-3B-Instruct adapted with completion-only LoRA supervised fine-tuning on 203 verified integer-answer mathematical rationales.

Training data

• 30 designated leaderboard-correction rationales from sft-30.jsonl
• 173 targeted Train rationales selected for similarity to the 831-problem n=11 majority-voting set
• Total: 203 unique examples
• Every completion was checked against its labeled integer and ends in \boxed{answer}
• Combined dataset SHA-256: b0d64ae9fff26f844056a1dae98e240a2a541889d379901dfac27c70a316c87c

The 173 targeted examples were generated without exposing the labeled answer to the model. The 30 correction examples are explicitly marked as calibration examples in the source dataset. No training examples are included in this model repository.

Training configuration

• Base revision: a1d308dfcc03e09da285d49d912439a655a571e8
• Objective: completion-only supervised loss
• LoRA rank: 16
• Learning rate: 1e-5
• Maximum sequence length: 2048
• Optimizer steps: 120
• Training split: all 203 examples; no held-out split
• Hardware: 2 x NVIDIA Tesla V100-SXM2 32GB

The uploaded weights are the merged Transformers checkpoint. This derivative remains subject to the Qwen Research License linked above.