Datasets:
Seb-9B (9B, fitted on train, one question per request): 0.754 acc / KL 0.349 / Brier 0.131 / ECE 0.031 on the official test split
Disclosure: I built Seb-9B (https://huggingface.co/ironbcc/seb-9b). These numbers are self-reported and have not been re-run by the maintainers.
Mode. Seb's training data included this benchmark's train split (none of the test cases), so it belongs in the "fitted or fine-tuned on train" table. It is a general model that takes other question schemas too, but on this benchmark it is not zero-shot.
Protocol deviation. Seb answers one question per request, so each case's five questions were scored as five calls with the same state. The README notes that request shape matters, so this is not strictly comparable with whole-case rows.
Results (all 400 cases, 2,000 decisions, 0 failures):
| Accuracy | KL from gold | Brier | ECE |
|---|---|---|---|
| 0.754 | 0.349 | 0.131 | 0.031 |
- By type:
noul0.862,choice0.753,score0.672. - Seb's accuracy is just above the 0.735 teacher self-agreement ceiling, which by the README's reading may partly reflect learning the teacher's quirks.
- Its KL is clearly worse than the fitted specialists' (about 0.08).
- Latency was not measured on this benchmark.
Metric definitions:
- Accuracy: argmax against the gold label.
- KL: KL(gold β prediction), natural log, mean per decision. The value doesn't depend on the floor used for near-zero probabilities, because Seb never outputs an exact zero.
- Brier: Ξ£(p β gold)Β² per decision, then the mean. The same code gives 0.1485 on our own whole-case Jev 1.13 run, against the published 0.148.
- ECE: 15 equal-width bins on top-1 confidence against the gold label. This is not the leaderboard's own ECE: the same code gives 0.046 for Jev, against the published 0.144.
I can share the per-decision probabilities if you'd like to rescore with your own scorer.