Large Language Model Evaluation on ARC, TruthfulQA, Winogrande, GSM8K, HellaSwag, and MMLU
73.7ARC AccuracyDNPO
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| DNPOBase Model=Zephyr-7B, Iteration=32025.02 | 73.7 | 41.7 | 76.6 | 33.6 | 82.7 | 58.6 | 61.2 | |
| DNPOBase Model=Zephyr-7B, Iteration=22025.02 | 73.5 | 39.7 | 76.5 | 32.3 | 82.8 | 58.7 | 60.6 | |
| DNPOBase Model=Zephyr-7B, Iteration=12025.02 | 73.4 | 38.1 | 76.6 | 33.4 | 82.7 | 58.3 | 60.4 | |
| SPINBase Model=Zephyr-7B, Iteration=02025.02 | 70.9 | 39.3 | 76.8 | 28.9 | 82.6 | 59 | 59.6 | |
| SPINBase Model=Zephyr-7B, Iteration=22025.02 | 70.7 | 37 | 76.1 | 27.6 | 82 | 58.5 | 58.6 | |
| SFTBase Model=Zephyr-7B, Iteration=SFT2025.02 | 70.4 | 34 | 76.2 | 31.8 | 81 | 58.8 | 58.7 | |
| SPINBase Model=Zephyr-7B, Iteration=32025.02 | 70.3 | 38.3 | 75.6 | 27.5 | 81.8 | 57.9 | 58.6 | |
| SPINBase Model=Zephyr-7B, Iteration=12025.02 | 70.2 | 36.2 | 76 | 31.6 | 81.7 | 58.5 | 59 | |
| DNPOBase Model=Llama-3.2-3B, Iteration=22025.02 | 64.6 | 33.8 | 72.1 | 30.3 | 57.7 | 56.6 | 52.5 | |
| DNPOBase Model=Llama-3.2-3B, Iteration=32025.02 | 64.6 | 33.7 | 73 | 30.7 | 58 | 56.8 | 52.8 | |
| DNPOBase Model=Llama-3.2-3B, Iteration=12025.02 | 64.3 | 33.7 | 72.4 | 29.6 | 57.5 | 56.5 | 52.3 | |
| SPINBase Model=Llama-3.2-3B, Iteration=32025.02 | 63.8 | 33.3 | 71.4 | 28.7 | 57 | 56 | 51.7 | |
| SPINBase Model=Llama-3.2-3B, Iteration=02025.02 | 63.7 | 33 | 71.8 | 28.4 | 56.9 | 56 | 51.6 | |
| SPINBase Model=Llama-3.2-3B, Iteration=12025.02 | 63.6 | 33.2 | 71.9 | 29.6 | 56.9 | 56 | 51.9 | |
| SPINBase Model=Llama-3.2-3B, Iteration=22025.02 | 63.5 | 33.2 | 72.1 | 29.1 | 56.9 | 56 | 51.8 | |
| SFTBase Model=Llama-3.2-3B, Iteration=SFT2025.02 | 63.4 | 32.6 | 72 | 27.8 | 56.8 | 56 | 51.4 |