Commonsense Reasoning on HellaSwag, Winogrande, and BoolQ (test)
70.6AccuracyMSFT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MSFTSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 70.6 | 4.12 | |
| IESSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 69.4 | 3.88 | |
| DynamixSFTSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 69 | 4.04 | |
| SFTSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 68.2 | 3.88 | |
| Continual SFTSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 65.2 | 1.71 | |
| BaseSize=Average, Evaluation Protocol=5-shot greedy decoding2026.03 | 18.4 | — |