~/system-one-bench / benchmarks/arabic-osact4-offensive
Text track · Arabic · All benchmarks
Offensive tweet, yes/no. Span-02 has the top score, 1.5 points ahead of d1, but d1 and 8 more are within sampling noise: a statistical tie for first.
| Rank | Model | Accuracy | 95% interval | Macro-F1 | Calibration error | Failed | |
|---|---|---|---|---|---|---|---|
| 1 | Span-02 RespanLeader | 97.0 | 94.6–99.4 | 95.1 | 0.9% | 0 | |
| 2 | d1 Liquid AITied for 1st | 95.5 | 92.6–98.4 | 92.8 | 5.1% | 0 | |
| 2 | 95.5 | 92.6–98.4 | 92.8 | 4.7% | 2 | ||
| 4 | 95.0 | 92.0–98.0 | 92.5 | 4.4% | 0 | ||
| 4 | GLiDE FastinoTied for 1st | 95.0 | 92.0–98.0 | 92.3 | 3.3% | 0 | |
| 6 | d1-ultra Liquid AITied for 1st | 94.5 | 91.3–97.7 | 91.5 | 5.4% | 0 | |
| 6 | 94.5 | 91.3–97.7 | 90.6 | 4.7% | 0 | ||
| 6 | Celeris-1 Decision CelerisTied for 1st | 94.5 | 91.3–97.7 | 91.1 | 5.1% | 0 | |
| 9 | 93.5 | 90.1–96.9 | 90.8 | 2.2% | 0 | ||
| 9 | 93.5 | 90.1–96.9 | 89.1 | 2.7% | 0 | ||
| 11 | d1.1 Liquid AI | 92.5 | 88.8–96.2 | 89.5 | 7.8% | 0 | |
| 12 | Mercury Decide Inception | 92.0 | 88.2–95.8 | 85.5 | 3.5% | 0 | |
| 13 | Mammoth 1 Baseten | 91.5 | 87.6–95.4 | 84.8 | 1.9% | 0 | |
| 13 | Dexy-Decision-9B Patrick-18 | 91.5 | 87.6–95.4 | 86.1 | 4.1% | 0 | |
| 15 | Elephant 1 Baseten | 91.0 | 87.0–95.0 | 84.4 | 3.5% | 0 | |
| 16 | 90.5 | 86.4–94.6 | 83.4 | 5.2% | 0 | ||
| 16 | 90.5 | 86.4–94.6 | 85.3 | 3.3% | 0 | ||
| 16 | 90.5 | 86.4–94.6 | 86.3 | 9.3% | 0 | ||
| 19 | Jev 1.13 TypeSafe | 90.0 | 85.8–94.2 | 86.3 | 5.5% | 0 | |
| 20 | 89.5 | 85.3–93.7 | 85.3 | 2.8% | 0 | ||
| 20 | 89.5 | 85.3–93.7 | 84.3 | 3.2% | 0 | ||
| 22 | laya-ara Wouzein-domain | 88.5 | 84.1–92.9 | 83.1 | 9.1% | 0 | |
| 22 | Kev 4B Jared Palmer | 88.5 | 84.1–92.9 | 83.4 | 6.8% | 0 | |
| 22 | 88.5 | 84.1–92.9 | 82.8 | 4.5% | 0 | ||
| 25 | Jev 1.14 TypeSafe | 87.0 | 82.3–91.7 | 82.9 | 4.8% | 0 | |
| 26 | 86.0 | 81.2–90.8 | 76.3 | 8.8% | 0 | ||
| 26 | 86.0 | 81.2–90.8 | 71.8 | 14.7% | 0 | ||
| 28 | laya-ara-rag Wouze | 84.5 | 79.5–89.5 | 69.2 | 5.0% | 0 | |
| 28 | 84.5 | 79.5–89.5 | 79.7 | 4.6% | 0 | ||
| 28 | 84.5 | 79.5–89.5 | 75.1 | 8.2% | 0 | ||
| 28 | d1-3B Liquid AI | 84.5 | 79.5–89.5 | 78.6 | 3.8% | 0 | |
| 32 | Joey 1 Baseten | 83.0 | 77.8–88.2 | 58.2 | 4.4% | 0 | |
| 33 | Laya Router Convai Innovations | 82.5 | 77.2–87.8 | 63.1 | 8.3% | 0 | |
| 33 | Naluri-279M Jericho Siahayaunusually high | 82.5 | 77.2–87.8 | 74.7 | 4.1% | 0 | |
| 35 | 80.5 | 75.0–86.0 | 67.3 | 5.1% | 0 | ||
| 36 | Classify-1 ScaleDown | 80.0 | 74.5–85.5 | 44.4 | 30.0% | 200 | |
| 37 | Naluri-308M Jericho Siahaya | 79.5 | 73.9–85.1 | 74.2 | 5.8% | 0 | |
| 38 | 79.0 | 73.4–84.6 | 73.0 | 3.9% | 0 | ||
| 39 | d1-omni-600M Liquid AI | 77.5 | 71.7–83.3 | 43.7 | 13.2% | 0 | |
| 40 | 72.5 | 66.3–78.7 | 64.9 | 2.9% | 0 | ||
| 41 | Kev 0.8B Jared Palmer | 49.0 | 42.1–55.9 | 48.0 | 15.1% | 0 |
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.