Datasets:
Amos 1 medium and large: general zero-shot scores, plus the large fitted on train
We'd like to add Amos 1 to the board. Two open-weight models scored on test, all 2,000 decisions, full distributions, through the same /v1/systemone API we serve publicly.
General, zero-shot (never trained on these four workflows, nor on train or test):
| Model | Accuracy | KL from gold | Brier | ECE | p50 latency |
|---|---|---|---|---|---|
| Amos 1 medium (9B) | 0.7645 | 0.151 | 0.080 | 0.103 | 118 ms‡ |
| Amos 1 large (27B) | 0.761 | 0.126 | 0.069 | 0.097 | 357 ms‡ |
General, then fitted on train (for your second table): the same large with one additive log-prior bias per question and option, fitted on 80% of train and checked on the other 20%. No weights change, and nothing is fitted on test.
| Model | Accuracy | KL from gold |
|---|---|---|
Amos 1 large, per-question priors fitted on train |
0.8025 | 0.122 |
‡ One H100, the client on the same machine, one request at a time. The large reads each choice question with its options in four rotated orders and averages them, which is why it is slower.
Metrics follow the card: accuracy is argmax against the gold label, KL is KL(gold‖pred) averaged over decisions, Brier is summed over options, ECE on the top answer.
Both models answer at https://api.amos.scira.ai/v1/systemone (no key) with "model": "amos-1-medium" or "model": "amos-1-large", so you can rescore them with the same client code as the Jev and Liquid rows. Weights are Apache 2.0, and the serving code is in each repo's serve/ folder. Happy to share the per-decision dumps.
