LLM Training Project - Usage Guide

Objective

Finetune the GPT2 model with data about "Jalal the cat", and then serve the new model and old models. The new fine-tuned model should correctly answer questions about Jalal the cat's space adventures, as mentioned in the training data. While the original GPT2 model should not be able to.

Environment Setup (One Time)

• Create a virtual environment: python3 -m venv llm_training_env

Environment Setup (Everytime)

• Activate the virtual environment: source llm_training_env/bin/activate
• Upgrade pip: pip install --upgrade pip
• Run the setup script: python scripts/setup.py (this installs all the dependencies from the requirements.txt file too)

Data

• Place your training data in: data/train_data.json
• Place your evaluation data in: data/eval_data.json

Configuration Files

• All layers fine-tuning config file: configs/config.yaml
• LoRA config file: configs/config_lora.yaml

Model Download

• To download the model, run: python scripts/download_model.py --config configs/training/config_lora.yaml

Device Setup

• To setup the device, run: python scripts/setup_device.py --config configs/training/config_lora.yaml

Full fine-tuning (All layers)

• To train the model, use: python scripts/train.py --config configs/training/config.yaml

LoRA Training

• To verify LoRA setup, run: python scripts/verify_lora_setup.py
• To train with LoRA, use: python scripts/train_lora.py --config configs/training/config_lora.yaml

List Models and Outputs

• To list all models or outputs, check the models/ and outputs/ directories directly.

Model serving

• Update the latest finetuned model directory in the serving config file.
• Start the server: python server.py
• Go to: http://0.0.0.0:8000/docs
• Try the two endpoints. Generate uses the new fine-tuned model, and Generate_base uses the original GPT2 model.