Language Modeling on Zero-shot Perplexity Suite (Lambada, Pile 10k, Wikitext)
33.69Average PerplexityBF16
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| BF16Model Size=13B, Precision=BF162025.01 | 33.69 | 12.42 | 22.45 | 75.06 | 24.81 | |
| FP4Model Size=13B, Precision=FP42025.01 | 33.99 | 13.67 | 21.62 | 75.84 | 24.83 | |
| BF16Model Size=7B, Precision=BF162025.01 | 35.06 | 14.34 | 23.33 | 77.72 | 24.86 | |
| FP4Model Size=7B, Precision=FP42025.01 | 35.62 | 14.29 | 24.42 | 78.42 | 25.36 | |
| FP4Model Size=1.3B, Precision=FP42025.01 | 36.86 | 15.33 | 23.07 | 82.52 | 26.51 | |
| BF16Model Size=1.3B, Precision=BF162025.01 | 37.38 | 14.98 | 25.1 | 82.77 | 26.65 |