L3.3-70B-Euryale-v2.3
Sao10K · 2024-12-07 · 70.6B parameters
L3.3-70B-Euryale-v2.3 is Sao10K's creative-writing and roleplay fine-tune of Meta's Llama 3.3 70B Instruct, released as a full BF16 checkpoint on Hugging Face under the repository's llama3 license tag. Its published Axolotl recipe uses rank-128 rsLoRA for one epoch at a 16,384-token sequence length with the paged_ademamix_8bit optimizer; the released checkpoint retains the base model's 131,072-token configuration.
Benchmark scores
| Benchmark | Score |
|---|---|
| SnakeBench Average Apples | 2.1 |
| SnakeBench Best Apples | 7.0 |
| SnakeBench Rating | 20.9 |
| SnakeBench Total Apples | 59.0 |
| SnakeBench Win Rate | 33.3 |