browsecomp-worker-sft-core-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).

This is the v1 (core-set-only) model: trained on the 541 demonstration trajectories where the MiniMax worker's own answer was judged correct. BrowseComp-Plus test set (n=150): zero-shot Qwen3-8B workers 0.513 → this model 0.620 → sibling mix model browsecomp-worker-sft-mix-v256 0.720 (= MiniMax-worker ceiling).

Kept public as (a) the ablation point showing what the +623 masked process-set trajectories add, and (b) the RL-stacking starting point (it has 10pt of headroom to the ceiling, unlike the mix model).

Training

541 core trajectories, full loss on assistant turns (thinking + tool calls + report), multi-turn qwen3 template, system/brief/tool observations masked. 3 epochs, lr 1e-5 cosine, 47 steps (miles sft_rollout). Harness (train = deploy): search snippets 256 words, ≤12 entries, worker budget 30720 tokens + budget wrap-up turn.

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