SFT fine-tune of Qwen/Qwen2.5-1.5B-Instruct on vukien2301/ultrainteract_math_cot_gpt54mini (78,349 GPT-5.4-mini regenerated math CoT problems from UltraInteract).

Training

Format: raw completion (no chat template) — "{question}\n{model_prediction}" + EOS
DoRA r=128, alpha=256, dropout=0.05
Target modules: q,k,v,o,gate,up,down
LR 2e-5, 1 epoch, batch 8 × grad-accum 4 (eff. 32)
Adapter merged into base via peft.merge_and_unload()
Hardware: 1× B200, bf16, flash_attention_2

Evaluation (lm-evaluation-harness, vLLM backend)

Task	Shots	Base Qwen2.5-1.5B-Instruct	This model
GSM8K (flex)	5	59.89	65.13
GSM8K (strict)	5	54.74	38.74
GSM-Plus (flex)	5	39.56	45.61
GSM-Plus (str)	5	32.45	24.18
MATH500	4	12.20	10.40

Strict-match drop is format drift (\boxed{} vs #### N), not capability loss — flex confirms.

Built for EMNLP H2H-SD compound-pipeline experiments. See [github / paper link TBD].