~/system-one-bench / benchmarks/banking77
Text track · English · All benchmarks
Banking intent, 77 options. Clef Flash has the top score, 2.5 points ahead of Clef Omni, but Clef Omni and 1 more are within sampling noise: a statistical tie for first. But its score is unusually high for Clef Flash: 98.5%, 10 points above the 88% its other datasets predict.
| Rank | Model | Accuracy | 95% interval | Macro-F1 | Calibration error | Failed | |
|---|---|---|---|---|---|---|---|
| 1 | 98.5 | 96.8–100.0 | 98.0 | 6.2% | 0 | ||
| 2 | 96.0 | 93.3–98.7 | 95.8 | 19.9% | 0 | ||
| 3 | 95.5 | 92.6–98.4 | 95.2 | 4.6% | 0 | ||
| 4 | Mammoth 1 Baseten | 94.5 | 91.3–97.7 | 94.2 | 2.8% | 0 | |
| 4 | Elephant 1 Baseten | 94.5 | 91.3–97.7 | 93.2 | 2.9% | 0 | |
| 6 | 93.0 | 89.5–96.5 | 89.3 | 5.0% | 0 | ||
| 6 | Joey 1 Baseten | 93.0 | 89.5–96.5 | 92.3 | 3.2% | 0 | |
| 6 | Dexy-Decision-9B Patrick-18in-domain | 93.0 | 89.5–96.5 | 92.6 | 3.1% | 0 | |
| 9 | GLiDE Fastino | 91.5 | 87.6–95.4 | 90.2 | 3.8% | 0 | |
| 10 | 91.0 | 87.0–95.0 | 90.0 | 7.1% | 0 | ||
| 11 | 90.5 | 86.4–94.6 | 89.6 | 4.7% | 0 | ||
| 11 | 90.5 | 86.4–94.6 | 88.9 | 4.4% | 0 | ||
| 13 | 90.0 | 85.8–94.2 | 86.5 | 1.8% | 0 | ||
| 14 | d1 Liquid AI | 89.0 | 84.7–93.3 | 86.1 | 3.6% | 0 | |
| 14 | Celeris-1 Decision Celeris | 89.0 | 84.7–93.3 | 85.5 | 3.3% | 0 | |
| 16 | Kev 4B Jared Palmerin-domain | 88.0 | 83.5–92.5 | 84.6 | 12.2% | 0 | |
| 16 | d1-3B Liquid AI | 88.0 | 83.5–92.5 | 84.7 | 3.1% | 0 | |
| 16 | 88.0 | 83.5–92.5 | 85.6 | 3.7% | 0 | ||
| 19 | 87.0 | 82.3–91.7 | 84.4 | 3.8% | 0 | ||
| 19 | 87.0 | 82.3–91.7 | 83.7 | 6.7% | 0 | ||
| 19 | d1.1 Liquid AI | 87.0 | 82.3–91.7 | 84.6 | 7.1% | 0 | |
| 22 | Mercury Decide Inception | 86.0 | 81.2–90.8 | 80.4 | 8.9% | 0 | |
| 23 | d1-omni-600M Liquid AI | 84.5 | 79.5–89.5 | 80.3 | 4.0% | 0 | |
| 24 | d1-ultra Liquid AI | 83.5 | 78.4–88.6 | 80.7 | 6.6% | 0 | |
| 25 | Span-02 Respan | 82.0 | 76.7–87.3 | 77.6 | 5.3% | 0 | |
| 26 | Kev 0.8B Jared Palmerin-domain | 80.5 | 75.0–86.0 | 76.0 | 11.4% | 0 | |
| 26 | 80.5 | 75.0–86.0 | 77.1 | 3.2% | 0 | ||
| 28 | 79.5 | 73.9–85.1 | 78.0 | 8.3% | 0 | ||
| 29 | Jev 1.13 TypeSafe | 79.0 | 73.4–84.6 | 75.1 | 10.1% | 0 | |
| 30 | Jev 1.14 TypeSafe | 78.5 | 72.8–84.2 | 76.6 | 10.4% | 0 | |
| 31 | 77.5 | 71.7–83.3 | 73.6 | 8.7% | 0 | ||
| 32 | Laya Router Convai Innovations | 44.0 | 37.1–50.9 | 35.8 | 49.0% | 0 | |
| 33 | Naluri-279M Jericho Siahaya | 31.0 | 24.6–37.4 | 24.6 | 11.1% | 0 | |
| 34 | Naluri-308M Jericho Siahaya | 27.0 | 20.8–33.2 | 21.3 | 9.5% | 0 | |
| 35 | 1.0 | 0.0–2.4 | 0.0 | 0.3% | 200 | ||
| 35 | 1.0 | 0.0–2.4 | 0.0 | 0.3% | 200 | ||
| 35 | 1.0 | 0.0–2.4 | 0.0 | 0.3% | 200 | ||
| 35 | Classify-1 ScaleDown | 1.0 | 0.0–2.4 | 0.0 | 0.3% | 200 | |
| 35 | 1.0 | 0.0–2.4 | 0.0 | 0.3% | 200 |
The bar is accuracy; the bracket on it is the 95% interval for 200 cases. Tied for 1st marks every model whose gap to the leader is within sampling noise (1.96 standard errors of the difference, treating the runs as independent). A failed call counts as wrong. In-domain models trained on another split of this dataset, by their maker's own account. Unusually high marks a score far above what that model's other datasets, and every other model's score here, predict (hover for the numbers). It often means the model trained on this dataset or one like it; it isn't proof, and the score still counts.
It counts towards the System One Index and its English sub-index. Every model gets the same cases and the same request, under the benchmark rules.