Commonsense Reasoning on Winogrande (NLL)
0.8NLLPPT (Distribution)
Evaluation Results
| Method | Links | |
|---|---|---|
| PPT (Distribution)Backbone=LLAMA3-8B, Temperature Scaling=false2026.05 | 0.8 | |
| PPT (Distribution) + TSBackbone=LLAMA3-8B, Temperature Scaling=true2026.05 | 0.8 | |
| Base + TSBackbone=LLAMA3-8B, Temperature Scaling=true2026.05 | 0.84 | |
| PPT (Mean)Backbone=LLAMA3-8B, Temperature Scaling=false2026.05 | 0.91 | |
| PPT (Distribution) + TSBackbone=QWEN2-7B, Temperature Scaling=true2026.05 | 0.96 | |
| Base + TSBackbone=QWEN2-7B, Temperature Scaling=true2026.05 | 1.03 | |
| BaseBackbone=LLAMA3-8B, Temperature Scaling=false2026.05 | 1.15 | |
| PPT (Distribution)Backbone=QWEN2-7B, Temperature Scaling=false2026.05 | 1.23 | |
| PPT (Mean)Backbone=QWEN2-7B, Temperature Scaling=false2026.05 | 1.88 | |
| BaseBackbone=QWEN2-7B, Temperature Scaling=false2026.05 | 2.18 |