Datasets:
Submission: FINAL-Bench/Darwin-27B-ZTC (general, zero-shot), accuracy 0.743
Hi! We would like to submit FINAL-Bench/Darwin-27B-ZTC for the zero-shot table.
Mode: general, zero-shot. The model never saw the Typed Decisions train split, its four workflows or its question schemas.
Model. Open weights (Apache-2.0), 27B, built on FINAL-Bench/Darwin-27B-RSI. It answers each typed question with one forward pass and a readout head over the options, so it generates no tokens. The inference code ships in the repo (autojev/, MIT).
Setup. test split, 400 cases, 2,000 decisions. One request per case with the state and all five questions, the same request shape as your rows. BF16 on one GPU. All 2,000 decisions answered, zero errors.
| Metric | Value |
|---|---|
| Accuracy | 0.743 |
| KL from gold | 0.204 |
| Brier | 0.097 |
| Type | Decisions | Accuracy |
|---|---|---|
| noul | 600 | 0.847 |
| choice | 600 | 0.723 |
| score | 800 | 0.675 |
Scoring. Accuracy is agreement with the gold label. KL is KL(gold || prediction), averaged over decisions. Brier is the squared error summed over the options of a decision, averaged over decisions. Our scorer reproduces your Uniform reference row (KL 0.444, Brier 0.238). We ran the test set twice; accuracy was 0.741 and 0.743, and the table reports the second run (the one with saved predictions). We did not measure latency in a setting comparable to your hosted rows, so we leave that column empty.
We are happy to share the per-decision predictions, or to help you re-run it. Thank you for the benchmark!
This is an official hf benchmark now, you can add the results to your model card and it should show up on the board, see - https://huggingface.co/docs/hub/en/eval-results
Correction to the per-type table above: it mixed two runs. The headline (accuracy 0.743, KL 0.204, Brier 0.097) is from our second run, the one with saved predictions, but the per-type accuracies came from the first run (0.741 overall). Here are the per-type numbers from the same second run:
| Type | Decisions | Accuracy |
|---|---|---|
| noul | 600 | 0.845 (507/600) |
| choice | 600 | 0.732 (439/600) |
| score | 800 | 0.675 (540/800) |
| all | 2,000 | 0.743 (1,486/2,000) |
The two runs differ on 14 of 2,000 decisions (9 better, 5 worse; mean +0.002, 95% case-bootstrap interval -0.002 to +0.006). The model card now reports every number from the second run. Thanks to the community member who spotted the mismatch.