Commonsense Reasoning on Winogrande (LIS, Accuracy (%))
3.4756LISDeepSeek V2 Lite
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeepSeek V2 LiteStrategy=Raw2026.02 | 3.4756 | 70.8 | |
| DeepSeek-MoEStrategy=Raw2026.02 | 3.1386 | 70 | |
| DeepSeek V2 LiteStrategy=R&Q2026.02 | 2.7207 | 71.4 | |
| DeepSeek-MoEStrategy=R&Q2026.02 | 2.4475 | 70.6 | |
| Switch Transformer (16 experts)Strategy=Raw2026.02 | 2.0063 | 49 | |
| Switch Transformer (16 experts)Strategy=R&Q2026.02 | 1.5394 | 48.5 | |
| Switch Transformer (8 experts)Strategy=Raw2026.02 | 1.5261 | 49.6 | |
| LLaMa-MoEStrategy=Raw2026.02 | 1.3925 | 67.3 | |
| LLaMa-MoEStrategy=R&Q2026.02 | 1.2791 | 66.7 | |
| Switch Transformer (8 experts)Strategy=R&Q2026.02 | 1.2146 | 49.2 |