Shangy/browsecomp-qwen3-32B-baseline-iter139

qwen3-32B-baseline-iter139

GRPO-trained deep-research agent on BrowseComp-Plus (search / open_page / finish tools over a fixed 100K-doc corpus). Cell qwen3-32B/baseline of the length-penalty experiment matrix, checkpoint at training iteration 139.

• Code: https://github.com/ys-2020/miles (branch browsecomp-rl, see docs/experiments/browsecomp-length-penalty-results.md for the full study)
• WandB: project browsecomp-b300
• Trained with miles fully-async GRPO: group size 8, global batch 256, temp 1.0, lr 1e-6, KL 0.001, ~100-turn ReAct rollouts, 40960-token context.

Offline eval (150 BrowseComp-Plus test questions, temp 0.6, 1 sample/q)

iter	accuracy	mean response len (tokens)	truncated ratio
19	0.293	2482	0.24
39	0.313	2586	0.19
59	0.347	2597	0.18
79	0.400	2527	0.13
99	0.407	2803	0.15
119	0.387	2820	0.13
139	0.427	3102	0.09

Resuming training in miles

Convert back with tools/convert_hf_to_torch_dist.py, rename the saved dir to iter_0000139, write 139 into latest_checkpointed_iteration.txt, then launch with RESUME=1 (see examples/browsecomp/slurm_gb300/).

training_metadata.json in this repo records provenance.