Shangy/browsecomp-qwen3-8B-async-iter179

qwen3-8B-async-iter179

GRPO-trained deep-research agent on BrowseComp-Plus (search / open_page / finish tools over a fixed 100K-doc corpus). Cell qwen3-8B/baseline of the length-penalty experiment matrix, checkpoint at training iteration 179.

• 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	0.273	2381	0.24
39	0.293	2584	0.26
59	0.313	2598	0.33
79	0.273	2683	0.30
99	0.367	2646	0.39
119	0.307	2856	0.43
139	0.353	2766	0.44
159	0.373	2751	0.39
179	0.360	2895	0.37

Resuming training in miles

Convert back with tools/convert_hf_to_torch_dist.py, rename the saved dir to iter_0000179, write 179 into latest_checkpointed_iteration.txt, then launch with RESUME=1 (see examples/browsecomp/slurm_gb300/).

training_metadata.json in this repo records provenance.