Language Modeling and Reasoning on ARC, BBH, HellaSwag, TruthfulQA, LAMBADA, WinoGrande, GSM8K, MT-Bench
54.61ARC (Accuracy)BitDelta
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| BitDeltaApplication=Parameter-efficient fine-tune (LoRA), Rank (r)=16, Source Model=Llama 2-7B UltraChat2024.02 | 54.61 | 34.28 | 79.1 | 46.6 | 70.58 | 69.3 | 15.16 | 52.8 | 4.87 | |
| Llama 2-7B UltraChatFine-tuning=r = 16 LoRA, Epochs=1, Training Dataset=UltraChat [17]2024.02 | 54.52 | 34.14 | 78.99 | 46.84 | 70.83 | 69.53 | 14.71 | 52.79 | 4.93 | |
| Llama 2-7BModel Type=Base Model2024.02 | 52.56 | 33.76 | 78.96 | 38.96 | 68.39 | 68.98 | 13.57 | 50.74 | — |