DVC and LLM Fine-Tuning

This project is an example about how DVC might be used for LLM tasks by use Openapi Api for fine tune a simple dataset.

Requirements

You need DVC cli installed on your environment to run the commands.

You will need credits on the Openapi API to get the key required on the fine_tune_model.py and predict.py scripts.

Usage

Install project required packages.

pip install -r requirements.txt

Add the dataset

dvc add data/dataset.csv

Create the stage prepare_dataset of the pipeline.

dvc stage add \
    -n prepare_dataset \
    -d data/dataset.csv \
    -d prepare_dataset.py \
    -d consts.py \
    -p prepare_dataset.count \
    -o data/train.jsonl 
    python prepare_dataset.py

Create the stage fine_tune_model of the pipeline.

dvc stage add \
    -n fine_tune_model \
    -d data/train.jsonl \
    -d fine_tune_model.py \
    -p fine_tune_model.n_epochs \
    -o result/new_model_name.txt \
    python fine_tune_model.py

Create the stage predict of the pipeline.

dvc stage add \
    -n predict \
    -p prepare_dataset.count \
    -d predict.py \
    -d result/new_model_name.txt \
    -d train.jsonl \
    -o result/predicted_dataset.csv \
    python predict.py

Create the stage evaluate of the pipeline.

dvc stage add \
    -n evaluate \
    -d result/predicted_dataset.csv \
    -d evaluate_predictions.py
    -o result/scores.json \
    python evaluate_predictions.py

At the end visualize your pipepiline by use the command dvc dag. And run all stages of the pipeline:

dvc repro

For troubloshooting the stages individualy you can run for example dvc repro prepare_dataset to run this stage only.