Natural Language Generation on Math
38.8BLEUADORE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ADOREBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 38.8 | 28.9 | 70.5 | |
| Full AttentionBackbone=Llama-2 7B, Context length=4,0002024.07 | 38.6 | 29.9 | 69.7 | |
| H2O(Rebuilt)Backbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 34.7 | 27.1 | 69.6 | |
| H2OBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 33.3 | 26.2 | 68.1 | |
| Strided AttentionBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 33 | 26.7 | 66.7 | |
| StreamingLLMBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 32.9 | 26.8 | 68.3 | |
| KV CompressionBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 32.2 | 24.4 | 66.4 | |
| Window AttentionBackbone=Llama-2 7B, Context length=4,000, KV cache size=1922024.07 | 30.3 | 24.3 | 66.3 |