~/system-one-bench / benchmarks/arabic-massive-scenario
Text track · Arabic · All benchmarks
Assistant request scenario, 18 options. Clef Flash and Dexy-Decision-9B share the top score, and 3 more are within sampling noise. But its score is unusually high for Clef Flash: 89.5%, 20 points above the 69% its other datasets predict.
| Rank | Model | Accuracy | 95% interval | Macro-F1 | Calibration error | Failed | |
|---|---|---|---|---|---|---|---|
| 1 | 89.5 | 85.3–93.7 | 82.6 | 4.9% | 0 | ||
| 1 | Dexy-Decision-9B Patrick-18Tied for 1stin-domainunusually high | 89.5 | 85.3–93.7 | 88.0 | 5.2% | 0 | |
| 3 | 89.0 | 84.7–93.3 | 84.8 | 6.3% | 0 | ||
| 4 | 88.0 | 83.5–92.5 | 83.1 | 7.5% | 0 | ||
| 5 | 83.5 | 78.4–88.6 | 77.3 | 7.0% | 0 | ||
| 6 | 79.0 | 73.4–84.6 | 74.2 | 5.3% | 0 | ||
| 7 | d1-ultra Liquid AI | 76.0 | 70.1–81.9 | 74.9 | 6.9% | 0 | |
| 8 | Mammoth 1 Baseten | 75.0 | 69.0–81.0 | 77.8 | 11.9% | 0 | |
| 9 | GLiDE Fastino | 73.5 | 67.4–79.6 | 68.1 | 13.6% | 0 | |
| 9 | Celeris-1 Decision Celeris | 73.5 | 67.4–79.6 | 66.0 | 6.3% | 0 | |
| 11 | d1 Liquid AI | 73.0 | 66.8–79.2 | 68.7 | 8.2% | 0 | |
| 11 | 73.0 | 66.8–79.2 | 72.0 | 11.8% | 0 | ||
| 11 | d1.1 Liquid AI | 73.0 | 66.8–79.2 | 72.7 | 10.0% | 0 | |
| 14 | 72.5 | 66.3–78.7 | 73.4 | 10.1% | 0 | ||
| 15 | Mercury Decide Inception | 72.0 | 65.8–78.2 | 74.4 | 16.5% | 0 | |
| 16 | d1-3B Liquid AI | 71.5 | 65.2–77.8 | 62.4 | 7.7% | 0 | |
| 16 | Joey 1 Baseten | 71.5 | 65.2–77.8 | 66.9 | 9.6% | 0 | |
| 18 | Span-02 Respan | 71.0 | 64.7–77.3 | 71.8 | 13.3% | 0 | |
| 19 | 70.5 | 64.2–76.8 | 64.1 | 8.3% | 0 | ||
| 20 | laya-ara Wouzein-domain | 69.5 | 63.1–75.9 | 64.6 | 21.8% | 0 | |
| 21 | Elephant 1 Baseten | 69.0 | 62.6–75.4 | 64.1 | 14.5% | 0 | |
| 22 | 66.0 | 59.4–72.6 | 64.3 | 16.8% | 0 | ||
| 23 | Jev 1.14 TypeSafe | 65.0 | 58.4–71.6 | 59.8 | 16.3% | 0 | |
| 24 | Naluri-308M Jericho Siahayain-domain | 63.5 | 56.8–70.2 | 55.3 | 11.1% | 0 | |
| 24 | 63.5 | 56.8–70.2 | 61.2 | 9.7% | 0 | ||
| 26 | Jev 1.13 TypeSafe | 63.0 | 56.3–69.7 | 57.3 | 21.0% | 0 | |
| 27 | 61.5 | 54.8–68.2 | 56.0 | 16.2% | 0 | ||
| 28 | laya-ara-rag Wouzein-domain | 60.5 | 53.7–67.3 | 55.6 | 22.3% | 0 | |
| 29 | 58.5 | 51.7–65.3 | 54.1 | 25.0% | 0 | ||
| 30 | 58.0 | 51.2–64.8 | 53.5 | 10.4% | 0 | ||
| 30 | Naluri-279M Jericho Siahayain-domainunusually high | 58.0 | 51.2–64.8 | 52.5 | 10.1% | 0 | |
| 32 | Kev 4B Jared Palmer | 57.0 | 50.1–63.9 | 54.7 | 12.9% | 0 | |
| 33 | 56.5 | 49.6–63.4 | 56.1 | 13.9% | 0 | ||
| 34 | 55.5 | 48.6–62.4 | 47.6 | 37.2% | 0 | ||
| 35 | 46.0 | 39.1–52.9 | 40.4 | 32.7% | 0 | ||
| 36 | 41.0 | 34.2–47.8 | 30.8 | 21.1% | 0 | ||
| 37 | Laya Router Convai Innovations | 39.0 | 32.2–45.8 | 37.3 | 40.6% | 0 | |
| 38 | Kev 0.8B Jared Palmer | 38.5 | 31.8–45.2 | 39.5 | 12.8% | 0 | |
| 39 | d1-omni-600M Liquid AI | 37.5 | 30.8–44.2 | 30.7 | 29.1% | 0 | |
| 40 | Classify-1 ScaleDown | 3.0 | 0.6–5.4 | 0.3 | 2.6% | 200 | |
| 40 | 3.0 | 0.6–5.4 | 0.3 | 2.6% | 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 Arabic sub-index. Every model gets the same cases and the same request, under the benchmark rules.