proposal-cot-sft Qwen2.5-32B LoRA (V2)
LoRA adapter for Qwen2.5-32B-Instruct fine-tuned on arxiv-proposal-cot-sft-32b-v2
(928 CoT examples, with_research_question conditioning, 16k sequence length).
- Recipe: LoRA r=64, α=128, all-linear; lr 1e-5 cosine (warmup 0.03), wd 0.01; effective batch 128 (8×H800); 1 epoch; bf16; verl FSDP SFT trainer. Train time ≈49 min.
- Behavior: generates
<thinking>+ a 10-tag XML research proposal from a reading-list prompt. Schema-complete and reference-grounded on held-out papers. - Honest result: on a 10-task implementation benchmark the SFT model matched but did not beat base pass-rate (base 1/10 vs SFT 0/10; several raw metrics improved). Released as a documented baseline for creativity-distillation research (the project's V3 direction).
- Apply with PEFT on top of
Qwen/Qwen2.5-32B-Instruct(Apache-2.0), or use the optional merged weights repo. Code + eval harness: https://github.com/XinghanLi66/autoresearch_idea_harness