browsecomp-worker-sft-mix-v256

Qwen3-8B fine-tuned to act as the search worker in an orchestrator-style multi-agent research system (frozen MiniMax-M2.7 coordinator delegates briefs; workers run multi-turn search/open_page/finish episodes over the BrowseComp-Plus corpus and return written reports).

Result: parity with MiniMax-M2.7 acting as the worker — BrowseComp-Plus test set (n=150, full set): zero-shot Qwen3-8B workers 0.513 → this model 0.720 = MiniMax-worker ceiling (0.720), at ~1/10 the inference cost.

Training

• Data: 1,164 worker trajectories demonstrated by MiniMax-M2.7 playing the worker role itself (bc_train, single pass, correct-episode filtered):
  • 541 core trajectories (worker's own answer judged correct) — full loss
  • 623 process trajectories (episode correct, worker's own conclusion wrong) — final report turn loss-masked, intermediate search behavior trained
• Loss: assistant turns only (thinking + tool calls + report), multi-turn qwen3 template; system/brief/tool observations masked. 3 epochs, lr 1e-5 cosine, ~107 steps (miles sft_rollout).
• Harness (train = deploy): search snippets 256 words, ≤12 entries/search, worker budget 30720 tokens with a budget wrap-up turn.

Code, pipeline and full experiment record: ys-2020/miles branch browsecomp-rl-b300, docs/experiments/orch-worker-sft-distillation.md.