MCQA Suite (ARC-Challenge, ARC-Easy, HellaSwag, Winogrande, SciQ, PIQA)
43.6ARC-Challenge AccuracyComposite - Stretched
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Composite - StretchedSize=8B2025.10 | 43.6 | 71.8 | 70.4 | 66.5 | 91.3 | 78.5 | 70.4 | |
| Composite - StackedSize=8B2025.10 | 42.7 | 71.5 | 70.6 | 64.8 | 91 | 77.4 | 69.7 | |
| Llama 3.2Size=8B2025.10 | 42.2 | 70.9 | 70.2 | 66 | 90.5 | 76.5 | 69.3 | |
| Composite - StackedSize=1B2025.10 | 33.5 | 59.5 | 56.7 | 59.4 | 83.2 | 73.1 | 60.9 | |
| Composite - StretchedSize=3B2025.10 | 33.5 | 66.9 | 57.9 | 57.5 | 84 | 72.3 | 62 | |
| Composite - StretchedSize=1B2025.10 | 32.5 | 61.5 | 56.1 | 56.1 | 88.1 | 72.5 | 61.1 | |
| Composite - StackedSize=3B2025.10 | 31.5 | 60.3 | 57.1 | 57.2 | 85.2 | 74.4 | 61 | |
| Composite - StretchedSize=350M2025.10 | 30.5 | 55.4 | 50 | 54.4 | 79.6 | 70.3 | 56.7 | |
| Llama 3.2Size=3B2025.10 | 30.1 | 58.6 | 54.3 | 56.4 | 83.1 | 72.8 | 59.2 | |
| Llama 3.2Size=1B2025.10 | 30 | 56.8 | 52.9 | 55.9 | 81.7 | 70.7 | 58 | |
| Composite - StackedSize=350M2025.10 | 27.9 | 53.9 | 49.6 | 54.8 | 78.7 | 71 | 56 | |
| Llama 3.2Size=350M2025.10 | 27.7 | 53.2 | 49.2 | 54.1 | 78.3 | 70.1 | 55.4 |