|
--- |
|
library_name: peft |
|
license: apache-2.0 |
|
language: |
|
- ru |
|
--- |
|
|
|
Use in the same way as IlyaGusev/saiga2_7b_lora. |
|
|
|
WARNING! Load tokenizer as AutoTokenizer.from_pretrained(model_path, use_fast=True) |
|
|
|
Up to 60% faster generation and 35% training (on identical russian text sequences!) with HF because of different tokenizer. |
|
|
|
Paper: Tikhomirov M., Chernyshev D. Impact of Tokenization on LLaMa Russian Adaptation //arXiv preprint arXiv:2312.02598. – 2023. |
|
|
|
## Training procedure |
|
|
|
ruadapt mistral trained on saiga corpuses. |
|
The quality is slightly worse than the IlyaGusev/saiga_mistral_7b_lora, but faster because of tokenizer. |
