vLLM Semantic Router · Open weights, local
d3 lite by vLLM Semantic Router scores 74.9 on the System One Index (15th of 24). Calibration error 7.1%.
| Dataset | Accuracy | Rank | Best on the board | Calibration error |
|---|---|---|---|---|
| AG News English | 85.5 | 21st of 24 | 92.5 | 10.1% |
| SST-2 English | 86.0 | 21st of 24 | 97.0 | 5.4% |
| SST-5 English | 50.5 | 16th of 24 | 62.5 | 4.9% |
| DAIR Emotion English | 57.0 | 12th of 24 | 91.5 | 18.2% |
| Banking77 English | 91.0 | 5th of 24 | 98.5 | 7.1% |
| BoolQ English | 81.0 | 19th of 24 | 92.5 | 8.5% |
| XNLI English English | 90.7 | 4th of 24 | 93.3 | 6.0% |
| Typed decisions English | 52.6 | 20th of 24 | 79.0 | 13.4% |
| XNLI French French | 81.3 | 9th of 24 | 90.7 | 4.7% |
| XNLI Arabic Arabic | 72.7 | 11th of 26 | 89.3 | 6.3% |
| MASSIVE scenario Arabic | 58.0 | 17th of 26 | 89.5 | 10.4% |
| MASSIVE intent Arabic | 77.5 | 13th of 26 | 91.5 | 3.4% |
| OSACT4 offensive Arabic | 80.5 | 21st of 26 | 95.5 | 5.1% |
| Tweet sentiment Arabic | 67.0 | 10th of 26 | 80.0 | 4.9% |
| MIRACL rerank Arabic | 82.5 | 14th of 26 | 90.5 | 4.3% |
| MIRACL pairwise Arabic | 84.5 | 18th of 26 | 91.0 | 5.7% |
vllm-sr/d3-liteDecision 3.0, text only (no mmproj). Apache 2.0.
Ranks count the models charted on the leaderboard. Every model gets the same cases and the same request, under the benchmark rules.