Hello,
I am encountering an error when prompting a fine-tuned model with a PDF for analysis, both in AI Studio and using the Gemini API. The entire chat session ceases to function after I prompt it with a document.
I would like to know if a Gemini fine-tuned model has document understanding capabilities. I have been unable to find this information in the official documentation.
Please advise.
Thank you,
Zak
Hi @zak,
Thanks for flagging. I am getting the same error in AI Studio. I will get back to you on this.
Hey @zak,
Currently,Tuning only support text input-output pair. Once tuned, it’s text input text output model. A tuned model doesn’t support image input. A pdf might contain image or you can say it will not support pdf as well, so that’s why you are getting this error.
Thanks.
Hello @Govind_Keshari,
Thank you for the clarification. I understand that the fine-tuned Gemini model only supports text input-output pairs.
Using AI Studio and the Gemini API, I have noticed the lack of support for a “system prompt” with fine-tuned models. While the documentation (Gemini Model Tuning Limitations ) mentions the absence of chat-style multi-turn conversations, it doesn’t explicitly address the use of a system prompt to guide the model’s behavior.
The absence of a system prompt would prevent the fine-tuned model from effectively utilizing RAG (Retrieval-Augmented Generation) or processing text extracted from PDF documents, as these techniques typically rely on providing contextual instructions via a system prompt.
Could you please confirm whether system prompts are indeed unsupported for fine-tuned Gemini models? If so, are there any alternative strategies recommended for incorporating contextual information or instructions when using RAG or PDF-derived text with a fine-tuned model?
Thanks,
Zak
Hey @zak,
Unfortunately, system instruction is also not supported with tuned model. It is also not mentioned in the doc. Here is the some limitations.
