Large Language Model Evaluation on MMLU, GSM8k, HellaSwag, and WinoGrande (test)
86.55MMLU AccuracyFP16
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FP16Format=-, Quantization=FP162025.09 | 86.55 | 95.07 | 86.22 | 84.93 | 88.19 | — | |
| GPTQ+Had128Format=NVFP, Quantization=GPTQ+Had1282025.09 | 85.59 | 94.16 | 85.56 | 84.77 | 87.52 | 99.24 | |
| GPTQ+Had16Format=NVFP, Quantization=GPTQ+Had162025.09 | 85.58 | 93.4 | 85.45 | 82.4 | 86.71 | 98.32 | |
| GPTQFormat=NVFP, Quantization=GPTQ2025.09 | 85.54 | 94.09 | 85.49 | 84.37 | 87.37 | 99.07 | |
| RTNFormat=NVFP, Quantization=RTN2025.09 | 85.5 | 93.48 | 85.63 | 83.27 | 86.97 | 98.61 | |
| RTN+Had128Format=NVFP, Quantization=RTN+Had1282025.09 | 85.24 | 91.81 | 84.91 | 83.35 | 86.33 | 97.89 | |
| RTN+Had16Format=NVFP, Quantization=RTN+Had162025.09 | 85.02 | 93.63 | 84.97 | 83.82 | 86.86 | 98.49 | |
| GPTQ+Had128Format=MXFP, Quantization=GPTQ+Had1282025.09 | 84.9 | 93.9 | 84.8 | 83.8 | 86.86 | 98.48 | |
| GPTQ+Had32Format=MXFP, Quantization=GPTQ+Had322025.09 | 84.82 | 94.54 | 84.66 | 83.11 | 86.78 | 98.4 | |
| RTN+Had128Format=MXFP, Quantization=RTN+Had1282025.09 | 84.37 | 94.47 | 84.22 | 82.4 | 86.37 | 97.93 | |
| RTN+Had32Format=MXFP, Quantization=RTN+Had322025.09 | 83.86 | 93.56 | 84.13 | 83.58 | 86.28 | 97.83 | |
| GPTQFormat=MXFP, Quantization=GPTQ2025.09 | 83.77 | 94.47 | 84.41 | 82.64 | 86.32 | 97.88 | |
| RTNFormat=MXFP, Quantization=RTN2025.09 | 83.42 | 92.65 | 83.93 | 81.45 | 85.36 | 96.79 |