Closed DylanDDeng closed 2 months ago
I guess the reason is that API service providers will add a conversation template for every query. That's why we need to generate instructions locally 😂.
Thanks for you explanation! That makes sense!
Let me know if you have further questions! I will close this issue for now.
Hi, I read your paper and your finding is really amazing and interesting. I try it using ollama with qwen2-7B-instruct and it really works. The diversity of the data is really nice. However, when I try this on the siliconflow which is a platform to provide some larger models like Qwen2-72B-Instruct. The model will only generate very official response like "hello what can I do for you? " .
I don't know why there exists huge differences of the data. Are something wrong for what I do?
Thanks for your work!