Summary of Finetuning vs Semantic Search
– Fine-tuning is a type of transfer learning that teaches a model a new task, but does not teach it new information.
– Semantic search allows searching based on meaning and context, not just keywords. It uses vector embeddings and scales well.
– Fine-tuning and semantic search are very different technologies. Fine-tuning is the wrong tool for doing QA on a corpus.
– Biggest misconception is that fine-tuning a model like GPT-3 on a new corpus will allow it to answer questions on that corpus.
– Fine-tuning only unfreezes a small part of the model. It does not retrain the whole model.
– Fine-tuning does not override hallucination or confabulation tendencies. The model can still make up answers.
– Fine-tuning is like transfer learning – you learn how to tie shoes, then you can tie other things. Not learning new info.
– So fine-tuning does not actually teach the model new information or content from your corpus.
– It just tweaks the model to be slightly better at a specific task, like QA. It does not add knowledge.