Few-shot Commonsense Reasoning on Commonsense Reasoning Suite (HellaSwag, PIQA, OBQA, COPA, WinoGrande)
27.3HellaSwag AccuracyD³
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| D³Granularity=Ours, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 27.3 | 55.7 | 15.2 | 63.5 | 51.1 | 42.56 | |
| Dynamic LossGranularity=Sample-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.6 | 56.8 | 13.8 | 59 | 50.1 | 41.26 | |
| Data Mixing LawGranularity=Domain-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.5 | 54.5 | 13 | 62 | 49.1 | 41.02 | |
| DoReMiGranularity=Domain-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.4 | 55.7 | 12.2 | 59 | 49.9 | 40.64 | |
| DoGEGranularity=Domain-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.2 | 55.8 | 11.5 | 62 | 50.4 | 41.18 | |
| Uniform SamplingGranularity=Sample-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.1 | 55.5 | 11.7 | 58 | 49.9 | 40.24 | |
| RegMixGranularity=Domain-level, Architecture=GPT-2 Medium (355M), Training Tokens=100B, Evaluation Protocol=Few-shot2026.05 | 26.1 | 55.6 | 13.2 | 60 | 50 | 40.98 |