lixiaochuan2020/acm-browsecompplus-qwen3.5-9b-opd-iter2

lixiaochuan2020/acm-browsecompplus-qwen3.5-9b-opd-iter2

Qwen3.5-9B fine-tuned by Offline On-Policy Distillation (OPD) — iteration 2 — for agentic deep research with a memory/context-management tool (MemTool regime), distilled from a Qwen3.5-397B-A17B teacher (top-20 forward-KL) on BrowseComp-Plus.

• —Base: Qwen/Qwen3.5-9B · Serving: vLLM, --tool-call-parser qwen3_xml, context window 131072.
• —BrowseComp-Plus eval (pass@1, MemTool 128K): anchor 63.5 → iter-1 67.7 → iter-2 69.8 → iter-3 72.7. This ckpt: iter-2 = 69.8%.

OPD training data (open-sourced)

• —Rollouts: lixiaochuan2020/acm-browsecompplus-train-rollouts-qwen3.5-9b-epoch2
• —Teacher top-20 logprob targets: lixiaochuan2020/acm-browsecompplus-teacher-logprobs-qwen3.5-9b-epoch2

Merged full-VLM weights (vision + LM), ready to serve directly.