physics-v07c-sft-qwen3.5-4b-merged
v07 SFT of Qwen/Qwen3.5-4B for EXACT-2026 Task-2 physics: emit a short (5-10 line)
reasoning preamble then ONE Python code block that computes FINAL ANSWER: / UNIT:.
Merged full model — serve directly with vLLM.
Training
- Method: 16-bit LoRA, train-on-completion (loss on the assistant turn only).
- LoRA: r=8, alpha=16, dropout=0.05, targets=['q_proj', 'k_proj', 'v_proj', 'o_proj'].
- Epochs=6, lr=3e-05, eff_batch=16, max_seq_len=2048.
- Data: 2634 train / 110 val trajectories (golden_60 held out).
- Chat template: Qwen
<|im_start|>/<|im_end|>; eos <|im_end|>, pad <|endoftext|>.
Metrics
- train_loss: 0.26994939584924715
- best_eval_loss: 0.26999393105506897
- best_eval_accuracy (val_56, selection metric): 0.7857142857142857
- (run
eval.py to fill)
External hint/teacher data used during data-gen is declared in the Data Disclosure Document.
