yuki-2025 / llama3-8b-fine-tuning-math

Fine-tuning Llama3 8b to generate JSON formats for arithmetic questions and process the output to perform calculations.

Fine-tuning Llama3 8b for Math

This project involves fine-tuning Llama3 8b to generate JSON formats for arithmetic questions and further post-process the output to perform calculations. This method incorporates the latest fine-tuning techniques such as Qlora, Unsloth, and PEFT. It enables faster training speeds and requires fewer computational resources.

PS: You need a T4 (16GB) GPU to run the code.

Colab Live code: https://drive.google.com/file/d/1NsSS1_M3pNAbkiBnPB3k5JKIkEQg3XNX/view?usp=sharing

Setup

1. Download all the files in this repo.
2. Run Load_model.py to load library and Llama3 8b.
3. Run Prepare_data.py to load function_call.jsonl dataset and prepare dataset.
4. Run Fine_tuning.py
5. Run Inference_n_save.py to test the fine-tuned models and save the model.

Preview

Before Fine-tuning:

Test1:

Test2:

Test3:

After Fine-tuning:

This is part of my research study in The University of Chicago. The data is come from: https://github.com/rohanbalkondekar/finetune_llama2