qweDante
Overview
Objective of this project is to produce a generative model able to reply to any questions in the style of Dante (sommo poeta). The chosen model is qwen3. The fine-tuning is done using Unsloth framework, to reduce the computation and memory footprint and running the fine-tuning on a free instance on Google Colab.
Data
The dataset has been created manually (gathering all the material by Dante Alighieri available online) and using Google Gemini 2.5 pro. Each line follows this pattern:
{"instruction": , "response": Training
The training is documented on the notebook present in the GitHub repo. I used a LoRA adapter to fine-tune model, loading the model in 4bit. Warmup steps are set to 5, and the batch size to 4. The total number of epoch was set to 1.
I built a small app on Google Colab. Try it.
