What is fine tuning?
The LLM can learn about item requirements more thoroughly through fine tuning the model, a process that involves adjusting the LLM’s internal weights. The weight adjustment customises the items that are generated to better match the syntactic features of items known to be effective. To fine tune the LLM, a dataset of prompt-item pairs is required as input, usually in JSON list format. Based on examples, autoregressive language modelling is used to identify adjusted weights that best predict the next token given all previous words. One drawback to keep in mind in generating items with fine tuned models is depending on the training data you may have less variability. However, the increased control it can provide often makes this worthwhile. The fine-tuned model can then be used to complete prompts. Here we will (soon) share code and results for fine-tuning the MiniLM sentence encoder.
Next section
Previous section
Return home
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).