Shangy/browsecomp-qwen3-8B-length-penalty-iter199

qwen3-8B-length-penalty-iter199

GRPO-trained deep-research agent on BrowseComp-Plus (search / open_page / finish tools over a fixed 100K-doc corpus). Cell qwen3-8B/length_penalty of the length-penalty experiment matrix, checkpoint at training iteration 199.

• 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	0.260	1973	0.25
39	0.240	2237	0.29
59	0.293	2080	0.30
79	0.280	2139	0.25
99	0.240	2156	0.31
119	0.300	2463	0.35
139	0.320	2521	0.36
159	0.353	2466	0.23
179	0.340	2563	0.19
199	0.373	2405	0.15

Resuming training in miles

Convert back with tools/convert_hf_to_torch_dist.py, rename the saved dir to iter_0000199, write 199 into latest_checkpointed_iteration.txt, then launch with RESUME=1 (see examples/browsecomp/slurm_gb300/).

training_metadata.json in this repo records provenance.