DHMahdi/LLM-Finetune

Finetuning an LLM on a Tunisian Arabizi sentiment analysis dataset # LLM-Finetune A complete solution for finetuning Mistral-7B on Tunisian Arabizi sentiment analysis tasks using Modal cloud infrastructure (You need a Modal account for access) and Axolotl. The project includes training pipeline, model serving, and a Streamlit interface for inference. # Prerequisites - Modal account and CLI installed - Hugging Face account and API token - Access to GPU resources (A100 or A10g) via Modal # Installation ## Set up modal - Grant modal access ``` python -m modal setup ``` - Create HuggingFace secret - Get access to use the model. ## Clone the repository ```Terminal git clone github.com ``` ```Terminal cd llm-finetune ``` ## Use a virtual environment (optional) ```Terminal python -m venv venv ``` - MacOS / Linux ```MacOS / Linux source venv/bin/activate ``` - Windows ```Windows venv\Scripts\activate ``` ## Install dependencies ```Terminal pip install -r requirements.txt ``` ## Start the training job ``` python -m modal run src.train --config=config/mistral7b.yml --data=datasets/data.jsonl ``` ## Serve the streamlit app for inference ``` python -m modal serve src/serve_streamlit.py ``` Or you can deploy if you're not making any changes ``` python -m modal deploy src/serve_streamlit.py ``` # Dataset Preparation Process This section describes the process of preparing and validating the sentiment analysis dataset. The process involves three main steps: converting CSV data to JSONL format, cleaning the dataset, and verifying the final data. ## Overview of Scripts 1. `csv_to_jsonl.py` - Converts the raw CSV dataset to JSONL format 2. `clean_dataset.py` - Cleans and validates the converted data 3. `verifydata.py` - Performs final verification of the prepared dataset ## Step 1: CSV to JSONL Conversion The `csv_to_jsonl.py` script converts the raw CSV dataset into JSONL format, which is more suitable for our sentiment analysis task. ```python # Example usage python csv_to_jsonl.py ``` K …