Zero-shot Evaluation on MMLU, HellaSwag, WinoGrande, PIQA, ARC-E, ARC-C, and OBQA
59.8MMLUDiLoCo K=1
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| DiLoCo K=1Model Size=15B, Optimizer=DiLoCo, K (Communication Frequency)=1, Zero-shot Evaluation Protocol=true2025.05 | 59.8 | 79.06 | 71.43 | 81.34 | 77.99 | 52.56 | 46.4 | 66.94 | |
| MuLoCo K=1Model Size=15B, Optimizer=MuLoCo, K (Communication Frequency)=1, Zero-shot Evaluation Protocol=true2025.05 | 59.49 | 79.06 | 71.11 | 81.61 | 78.11 | 52.73 | 46.6 | 66.96 | |
| DP MuonModel Size=15B, Optimizer=DP Muon, Zero-shot Evaluation Protocol=true2025.05 | 59.14 | 79.05 | 71.19 | 81.12 | 77.23 | 53.07 | 46.8 | 66.8 | |
| DP AdamWModel Size=15B, Optimizer=DP AdamW, Zero-shot Evaluation Protocol=true2025.05 | 57.2 | 77.69 | 71.19 | 81.39 | 78.11 | 52.3 | 45.4 | 66.18 | |
| MuLoCo K=16Model Size=15B, Optimizer=MuLoCo, K (Communication Frequency)=16, Zero-shot Evaluation Protocol=true2025.05 | 56.96 | 77.18 | 70.72 | 80.74 | 76.6 | 53.16 | 46.2 | 65.94 | |
| DiLoCo K=16Model Size=15B, Optimizer=DiLoCo, K (Communication Frequency)=16, Zero-shot Evaluation Protocol=true2025.05 | 54.15 | 77.59 | 69.3 | 81.72 | 77.65 | 52.56 | 46.8 | 65.68 |