Quelles sont les différentes techniques de fine-tuning d’un modèle ?

Leur avantage et inconvénients ? L’exemple de Flant T-5

In context learning (ICL)

• Zero shot
• One shot
• Few shot learnings

Drawbacks

• For small models it won’t work
• Limited by the context size

Fine tuning is the process of giving a structure dataset with examples and labels to updates the wieghts of the LLM

Labels examples are prompt-completion pairs

Prompts fine tuning

Summarize the following text:
[Text to summarize]
[Summary]

jinja: "Given the following review: \\n {{review_body}} 
\\n predict the associated rating from 1 to 5 

Prepare training data

Divide into training, validation, and test set

This is an instructed LLM