Qwen3-1.7B browsing worker — SFT iteration 1366
Evaluation-ready full BF16 export of qwen1p7b-core541-gate50-sft-0916/hf/iter_1366_eval. This is the SFT checkpoint evaluated under six orchestrators on BrowseComp and DR9K, not an RL checkpoint. Training mixture/optimizer details are not independently reconstructed here; the run name is provenance, not a validated data description.
Base: Qwen/Qwen3-1.7B revision 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e. Config context is 40,960 tokens. The evaluation export trims padded rows of model.embed_tokens.weight and lm_head.weight to the tokenizer vocabulary; see eval-preparation.json and payload-verification.json. All 311 tensors were validated. Tokenizer files match the pinned raw checkpoint. Uploaded weight bytes exactly match the evaluated export; see SHA256SUMS.
Evaluation
Worker under orchestrator, concurrency 64; BrowseComp uses four TP1 replicas, DR9K uses TP1 native DP4. BrowseComp has 150 questions (50 per difficulty); DR9K has 256 (128 frontier + 128 challenge). This is not solo-model accuracy. Repository manual protocol uses normalized exact matching, identical-answer precedents and full-answer agent review, not the canonical Gemini judge. Empty and failed episodes count wrong.
| Orchestrator | BrowseComp | DR9K |
|---|---|---|
| MiniMax-M2.7 | 103/150 | 112/256 |
| DeepSeek-V4-Flash | 99/150 | 152/256 |
| Nemotron Ultra | 99/150 | 137/256 |
| MiMo-V2.5 | 102/150 | 149/256 |
| Nemotron Super | 96/150 | 128/256 |
| Inkling-Small | 106/150 | 159/256 |
Full raw/SFT report includes exact latency settings, difficulty scores, role token counts and measured versus theoretical prefix reuse. BrowseComp traces and DR9K traces.
Use the per-benchmark chat templates and evaluator in the report artifacts to reproduce worker behavior. Generic chat generation does not reproduce the search tools, budgets or orchestrator setup.
Grading erratum (2026-09-18): MiMo and Inkling DR9K each gain one answer under the existing accepted Bretton Woods Agreement alias rule. Predictions and weights unchanged.