BiLLa is an open-source reasoning-enhanced bilingual LLaMA model in Chinese and English. The main features of the model are:
- Greatly improve the Chinese comprehension ability of LLaMA, and minimize the damage to the original LLaMA English ability;
- Increase more task-type data during training , use ChatGPT to generate analysis, strengthen the model understanding task solving logic;
- Full parameter update, pursue better generation effect.
Due to limited personal energy, I was unable to fully compare and evaluate BiLLa with the current mainstream open source models. The following are the conclusions drawn after limited evaluation and analysis:
- BiLLa-7B-LLM is significantly better than Chinese-LLaMA-7B in Chinese-English language modeling ability;
- BiLLa-7B - SFT Chinese reasoning ability is significantly better than BELLE-LLaMA-Ext-7B and other models;
- Scored by GPT4, BiLLa-7B-SFT scores significantly higher than ChatGLM-6B in English instructions, and the Chinese score is the same, But problem solving and code score higher.
Visit Official Website