Fine-tuning an LLM involves taking an existing, pre-trained model and doing some light retraining/tuning of the parameters so that the model performs better at a given, specific task.
Fine-tuning is just one strategy for altering the behavior of LLMs. Other strategies include Retrieval Augmented Generation and creating AI Agents.
Fine-tuning might be the best option for augmenting LLMs in the following cases:
- We need the LLM to adopt a very specific style when generating text output
- The output needs to have a very specific format/structure
- Reduced latency or offline use. Unlike RAG or AI Agents, fine-tuned models don’t need to call any databases or external services.
- Learning complex rules. If we need the model to understand intricate physical or chemical processes, for instance.
Example
See this Kaggle notebook for an example of fine-tuning the Gemini flash 1.5 model to better classify text data.