adarsh k.

Hello, Following the fine-tuning of the embedding model, are there any visual (like retrieved documents are closer to the question in 2d or something similar) comparison methods or code snippets available to assess how the fine-tuned model's embeddings better represent data, particularly in the context of RAG, in contrast to the embeddings generated by the base model?

Hi Amber R. As you previously noted, we initially began our work with RAG and prompt engineering, but unfortunately, the results were not as promising as we had hoped. As a result, we decided to fine-tune the embedding models using the MNR loss, and we observed a significant improvement in the hit rate. Now, we are considering fine-tuning a generator model. However, it's worth noting that we currently only have access to question and passage pairs; we do not have answer data. Consequently, we are actively exploring methods to fine-tune the generator model using only these question and passage pairs.

Hi Team, I had a question regarding fine-tuning the BART or any generator model for question answering tasks. In (Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks paper)they use a dataset with question-and-answer pairs for fine-tuning the BART generator model(correct me if I am wrong here). and All the resources I went through had question-and-answer pairs for fine tuning. If I only have data with questions and their corresponding passages, how can I fine-tune the generator model to improve answers for my custom data? I would greatly appreciate your guidance on this. Looking forward to your response.