The model is derived from Llama-2-7b-hf through pruning using LLM-Streamline (Streamlining Redundant Layers to Compress Large Language Models, ICLR 2025 Spotlight). The entire training process required only 0.06B tokens.
Below are the results of the evaluation using lm-eval:
| arc_c | arc_e | boolq | hellaswag | openbookqa | rte | winogrande | Avg |
|---|
| Llama-2-7B | 43.3 | 76.4 | 77.7 | 57.2 | 31.4 | 62.8 | 69.1 | 59.7 |
| Llama-2-4.7B | 34.0 | 64.6 | 74.7 | 49.8 | 27.4 | 61.7 | 66.4 | 54.1 |
