Qwen3-4B BrowseComp Worker SFT (0915)
Final full-parameter SFT worker checkpoint iter_1366, trained on four NVIDIA GB300 GPUs using Miles. No RL has been applied. Intended as a research worker under a frozen, self-hosted MiniMax-M2.7 coordinator.
Code and reproduction: Miles, paper/browsecomp-4b-training-nvidia-b300-0915.
Evaluation
One full round on the same frozen 150-question BrowseComp-Plus/FoldAgent test split, using MiniMax-M2.7 coordinator, hybrid BM25+dense RRF retrieval and Gemini 3.6 Flash grading. Runtime STOP gate is off.
| Worker | Easy | Medium | Hard | Total |
|---|---|---|---|---|
| Original Qwen3-4B | 43/50 | 29/50 | 11/50 | 83/150 (55.33%) |
| This SFT model | 49/50 | 41/50 | 18/50 | 108/150 (72.00%) |
Paired improvement: 31 questions improved, 6 regressed, +16.67 percentage points. SFT accuracy Wilson 95% interval: 64.33–78.57%. These are system scores with the coordinator and retrieval, not standalone model scores. This is a single stochastic round, not pass@k. Budget-exhausted episodes remain in the denominator.
Training
- Initialization:
Qwen/Qwen3-4Brevision1cfa9a7208912126459214e8b04321603b3df60c(pretrained/instruction model, not random initialization). - Core:
Shangy/browsecomp-worker-sftrevisionf1f0fb3053ef0cd15831a58e7943901340666630,data/core.jsonl, 541 trajectories repeated 3 times. - Gate:
Shangy/browsecomp-worker-gate-sft-v1revisioneb62ff5f74592b3409488fb52281d94ab3c5f22d,gating/train.jsonl, 18,698 source rows (already includes hard-negative repeats). - Balance by exact Miles Qwen3 supervised token masks: 1,623 core rows + 42,137 sampled gate rows = 43,760 rows, 3,952,827 worker vs 3,952,818 gate supervised tokens. Seed 42.
- No 904-row worker replay, no
mix.jsonl, no gate validation/holdout included. - Adam, LR 1e-5, cosine to 1e-6, warmup 10%, weight decay 0.1, beta=(0.9,0.95), batch 32, TP4, context 40,960, full recomputation, one constructed-mixture epoch.
- Miles uses floor division for epoch length: 1,367 rollout slots; release-checkpoint initialization starts at rollout 1, ending at 1,366. See run metadata for actual updates.
- Training template:
miles/utils/chat_template_utils/templates/qwen3_fixed.jinja; assistant-only loss.
Serving
Use the same evaluation keep-thinking template as the raw baseline:
examples/browsecomp/gb300_main_eval_0907/templates/qwen3_fixed_keepthink.jinja in the linked repository.
SGLang: TP1 per worker, context 40,960; worker episode budget 24,576 tokens, up to 40 turns, max 4,096 generated tokens per turn, temperature 1, top-p .95.
The model emits the BrowseComp XML-style search, open_page, and finish calls; use the repository's worker prompts and environment.
This repository contains the exact serving variant used for the SFT evaluation: 128 vocabulary-padding rows removed from the exported embedding, all original vocabulary rows retained byte-for-byte. Tokenizer files are restored from the pinned original Qwen3-4B; export reserialization changed defaults but preserved vocabulary, merges and added tokens. The original training export is retained in the experiment workspace.
Limitations
Core exports lack parent-question IDs, so question-level test exclusion cannot be independently proven from that file. Gate question hashes were checked: all originate from the 680-question training split, with disjoint train/validation/holdout and no overlap with the 150 test questions. This checkpoint should not be described as a proven contamination-free benchmark model.
The documented historical 8B SFT mixture was adapted to 4B with exact current loss-mask token balancing; it is not a byte-identical reproduction of the historical mixer. Model weight-content hashes are published in SHA256SUMS.