PengQu/Llama-2-7b-vicuna-Chinese
PengQu/Llama-2-7b-vicuna-Chinese is a 7 billion parameter chat model, fine-tuned on the Llama-2-7b-hf foundation model, specifically optimized for enhanced performance in both English and Chinese. It leverages a blend of ShareGPT and ShareGPT-ZH datasets, demonstrating improved MMLU and C-Eval scores compared to its base Llama-2 and Vicuna 1.1 counterparts. This model is designed for conversational AI applications requiring robust bilingual capabilities and a less cautious response style than the original Llama2-chat.
Loading preview...
Overview
PengQu/Llama-2-7b-vicuna-Chinese is a 7 billion parameter conversational model built upon the commercially available meta-llama/Llama-2-7b-hf foundation model. It has been supervised fine-tuned using a combination of English and Chinese ShareGPT data, including ShareGPT, ShareGPT-ZH, and Langchain-MRKL-finetune datasets, with training code based on FastChat.
Key Capabilities & Improvements
- Enhanced Bilingual Performance: Demonstrates improved capabilities in both English and Chinese compared to the original Llama-2 and Vicuna 1.1 models.
- English Evaluation (MMLU): Achieves an MMLU score of 48.8, outperforming Llama-2-7b (45.3) and Vicuna 1.1 (44.8).
- Chinese Evaluation (C-Eval): Scores 34.7 on C-Eval, surpassing Llama-2-7b-chat (30.3) and Vicuna 1.1 (30.3).
- Less Cautious Responses: Empirical results indicate that the model avoids the overly cautious response style sometimes observed in Llama2-chat.
Use Cases
This model is particularly well-suited for applications requiring a strong bilingual conversational agent, especially where improved performance in both English and Chinese is critical. Its fine-tuning on diverse ShareGPT data makes it effective for general-purpose chat and dialogue systems.