When is fine-tuning necessary?
As we saw in previous chapters, good prompt engineering combined with the non-parametric knowledge you can add to your model via embeddings are exceptional techniques to customize your LLM, and they can account for around 90% of use cases. However, the preceding affirmation tends to hold for the state-of-the-art models, such as GPT-4, Llama 2, and PaLM 2. As discussed, those models have a huge number of parameters that make them heavy, hence the need for computational power; plus, they might be proprietary and subject to a pay-per-use cost.
Henceforth, fine-tuning might also be useful when you want to leverage a light and free-of-charge LLM, such as the Falcon LLM 7B, yet you want it to perform as well as a SOTA model in your specific task.
Some examples of when fine-tuning might be necessary are:
- When you want to use an LLM for sentiment analysis on movie reviews, but the LLM was pretrained on Wikipedia articles and books. Fine-tuning...
