Spaces:
Running
Running
Why does llama-3-70b on hf.co/chat have excellent Chinese capabilities?
#496
by
wencan-lab
- opened
I have tested multiple model services of llama-3-70b, and their Chinese capabilities are poor. The Chinese Internet also unanimously believes that the Chinese capabilities of llama-3 cannot meet the requirements.
But today I found that llama-3-70b on hf.co/chat has excellent Chinese capabilities, and its performance is not inferior to that of other open-source models.
Is it because hf used an optimized version of llama-3, or did meta update the parameters of llama-3?
IIRC we use full precision for our models on HuggingChat whenever possible, so maybe other apps you tried llama3 on used a quant which could affect performance ? Not sure, it's always hard to evaluate performance without a proper benchmark 😅