Seeking Advice on Fine-Tuning LLMs for Generating Documents

Beginners
1

Hello everyone,

I hope this message finds you well. I am currently working on a project that involves generating technical documents, specifically CCTP (Cahier des Clauses Techniques Particulières), from DQE ( a document which contains the summary for the drafting of the cctp) using large language models (LLMs). I have access to several examples of both CCTP and DQE, as well as a powerful GPU setup with an A100 40GB.

I am looking for advice on the following aspects:

  1. Model Selection: Which open-source LLMs would be best suited for fine-tuning to generate detailed technical documents? I am considering models like BLOOM, T5, BART, Llama, and Mistral.
  2. Data Preparation: How should I structure and prepare my training data to effectively fine-tune these models? I have extracted text from PDFs and need guidance on annotation and creating input-output pairs.
  3. Fine-Tuning Process: Any tips or best practices for fine-tuning these models on my specific task? I am particularly interested in ensuring the generated documents are accurate and coherent.

I would greatly appreciate any insights, resources, or experiences shared by the community. Thank you in advance for your help!

Best regards,

2

If it’s an error or something, we can deal with it to a certain extent on this forum. However, I think it’s more reliable to ask about specialized topics such as LLM tuning or training know-how for generative AI on HF Discord.