Datasets:
Submission: Qwen/Qwen3.8-27B base model (general, zero-shot): 0.661 acc / 0.590 KL / 0.214 Brier, plus FP8, NVFP4 and Swift 1.5
The board has Qwen3.8-27B fine-tunes (jev-vision-27b, Qwen3.8-27B + LoRA, 0.714) but no row for the plain base model. Here it is, so those rows have a reference. I'm an outside evaluator, not the model authors.
| Model | Mode | Acc | KL | Brier | ECE | p50 per case |
|---|---|---|---|---|---|---|
| Qwen/Qwen3.8-27B (BF16, rev 1d4bf0f2) | general, zero-shot, open weights | 0.661 | 0.590 | 0.214 | 0.143 | 381 ms |
By type (BF16, accuracy / KL): choice 0.647 / 0.576, noul 0.697 / 0.509, score 0.645 / 0.662.
Full test prediction trace (400 cases, every distribution): results/stock-bf16/decisions/pass1.jsonl.
Quantization leaves accuracy within 0.005: Qwen/Qwen3.8-27B-FP8 (017b9c7a) 0.663 acc / 0.558 KL, RadixArk/Qwen3.8-27B-NVFP4 (319f741c) 0.666 / 0.542. Swift 1.5 NVFP4 (ukisai/Swift-1.5-Qwen3.8-27b-NVFP4, 25482027), a general chat fine-tune of the same base: 0.683 acc / 0.676 KL / 0.216 Brier / 0.140 ECE.
How it was scored: test split (400 cases, 2,000 decisions), dataset revision 1415d375 (the all/test parquet is identical at the current head), the dataset's score_one. One request per case with the state and all five questions, concurrency 1. Served with SGLang (lmsysorg/sglang:nightly-dev-cu13-20261010-961bcf48) on /v1/systemone, not the card's vLLM harness, thinking off, one RTX PRO 6000 at 400 W.
SGLang caveat: with this model a warm prefix cache changes answers. On the first 50 test cases, 25 of 250 top answers (NVFP4) differ between a cold and a warm pass of the same prompt. The numbers above are mostly cold-cache.
For the fitted table, if useful: fitted on train and applied to the returned probabilities in the client (no weight or server change). T is one temperature per question type; T+B adds one log-prior bias per workflow, question and option (same idea as the Amos 1 large priors row).
| Model | Fit | Acc | KL | Brier | ECE |
|---|---|---|---|---|---|
| RadixArk/Qwen3.8-27B-NVFP4 | T | 0.666 | 0.203 | 0.105 | 0.084 |
| RadixArk/Qwen3.8-27B-NVFP4 | T+B | 0.743 | 0.133 | 0.069 | 0.135 |
0.743 is only about 0.01 above the card's 0.735 teacher self-agreement, so I read it as fitting the teacher, not as more correct.
All rows, paired intervals, per-type tables, noise checks and the rerun command: https://github.com/adcostanza/benchmarks/tree/main/qwen38-27b-decisions
Thanks, this is a useful reference row. Your pass1.jsonl rescored to the same numbers (0.661 accuracy, KL 0.590, Brier 0.214), so the Qwen3.8-27B base model is now in the zero-shot table at #31 (PR #63), listed as an outside evaluation.
I left the FP8, NVFP4 and Swift runs and the client-side T and T+B fits out of the table for now; the base row is the reference the Qwen3.8 fine-tunes were missing. The cold versus warm prefix-cache difference on SGLang is worth knowing.