Shangy/browsecomp-qwen3-8B-length-penalty-global-ref-iter259

qwen3-8B-length-penalty-global-ref-iter259

GRPO-trained deep-research agent on BrowseComp-Plus (search / open_page / finish tools over a fixed 100K-doc corpus). Cell qwen3-8B/length_penalty_global_ref of the length-penalty experiment matrix, checkpoint at training iteration 259.

• 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	0.247	2334	0.22
39	0.240	2244	0.20
59	0.260	2325	0.21
79	0.280	2285	0.13
99	0.300	2479	0.17
119	0.273	2541	0.17
139	0.333	2280	0.14
159	0.340	2339	0.07
179	0.373	2143	0.06
199	0.340	2026	0.05
219	0.360	1986	0.02
239	0.407	1988	0.01

Resuming training in miles

Convert back with tools/convert_hf_to_torch_dist.py, rename the saved dir to iter_0000259, write 259 into latest_checkpointed_iteration.txt, then launch with RESUME=1 (see examples/browsecomp/slurm_gb300/).

training_metadata.json in this repo records provenance.