OFA-Sys / InsTag

InsTag: A Tool for Data Analysis in LLM Supervised Fine-tuning
196 stars 7 forks source link

Details about Qwen-1_8B Instagger. #10

Open litsh opened 1 month ago

litsh commented 1 month ago

I see there is a Qwen-1_8B version of Instagger on ModelScope. Could you please share the prompt you used for finetuning this model so that we can obtain better results when using the tagger.

Lukeming-tsinghua commented 1 month ago

Hi litsh, the tagger is trained with vanilla prompts and a default system prompt "You are a helpful assistant.", so no other template prompts are used for inference. For example, if you would like to tag "How are you?" then just use it as the query. Please kindly note that this is a Qwen model finetuned with the chatml format like other models in the Qwen series. So please make sure you generate prompt with proper chatml template.

litsh commented 1 month ago

Hi litsh, the tagger is trained with vanilla prompts and a default system prompt "You are a helpful assistant.", so no other template prompts are used for inference. For example, if you would like to tag "How are you?" then just use it as the query. Please kindly note that this is a Qwen model finetuned with the chatml format like other models in the Qwen series. So please make sure you generate prompt with proper chatml template.

Thanks for you reply! If I want to tag an instruction "How are you?", the full prompt feed into the model should belike which one of the below: <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user How are you?<|im_end|> <|im_start|>assistant or <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user Please identify tags of user intentions in the following user query and provide an explanation for each tag. Please respond in the JSON format {"tag": str, "explanation": str}. Query: How are you? <|im_end|> <|im_start|>assistant

SefaZeng commented 1 month ago

Hi litsh, the tagger is trained with vanilla prompts and a default system prompt "You are a helpful assistant.", so no other template prompts are used for inference. For example, if you would like to tag "How are you?" then just use it as the query. Please kindly note that this is a Qwen model finetuned with the chatml format like other models in the Qwen series. So please make sure you generate prompt with proper chatml template.

Thanks for you reply! If I want to tag an instruction "How are you?", the full prompt feed into the model should belike which one of the below: <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user How are you?<|im_end|> <|im_start|>assistant or <|im_start|>system You are a helpful assistant.<|im_end|> <|im_start|>user Please identify tags of user intentions in the following user query and provide an explanation for each tag. Please respond in the JSON format {"tag": str, "explanation": str}. Query: How are you? <|im_end|> <|im_start|>assistant

Hi, have you figured out this?

litsh commented 1 month ago

at {"tag": str, "explanation": str}. Query: How are you? <|im_end|> <|im_start|>assistant

Hi, have you figured out this? @SefaZeng Use the former one should be fine.