LLM Training (Refactored)

This repository provides a modular, extensible framework for data preparation, fine-tuning, evaluation, and compression of large language models (LLMs).

Key Features

• Modular Python package: llm_training
• Unified CLI: llm-train for data prep, training, and evaluation
• Configuration via JSON/YAML files with OmegaConf
• Support for tensorized model compression (Llama, Mixtral)
• Integrated WandB logging for experiment tracking
• Unit tests with pytest
• MIT licensed with community guidelines

Quickstart

1. Install dependencies:

   pip install -e .

2. Run data preparation:

   llm-train data-prep --config configs/data_prep_config.json

3. Fine-tune a model:

   llm-train train --config configs/sft_config.json

4. Evaluate a model:

   llm-train eval --config configs/evaluate_config.json

For model compression, see the scripts in scripts/compression/.

Configuration

All workflows are configured via JSON or YAML files using OmegaConf. See configs/ for examples. The config files specify model paths, hyperparameters, dataset paths, and training arguments.

Example Configs

• data_prep_config.json - Dataset preparation configuration
• sft_config.json - Supervised fine-tuning parameters
• evaluate_config.json - Evaluation benchmark settings
• accelerate_config.yaml - Multi-GPU training with Accelerate

Project Structure

• llm_training/ - Main Python package with core functionality
• scripts/ - Standalone scripts for data prep and compression workflows
• configs/ - Configuration files for different workflows
• tests/ - Unit tests

Testing

Run all tests with:

pytest

Docker

Build and run in a reproducible environment:

docker build -t llm_training .
docker run -it llm_training

Contributing

See CONTRIBUTING.md for guidelines. All contributions and issues are welcome!

Changelog

See CHANGELOG.md for release history.

Citation

If you find this work useful, please cite it as follows:

@misc{sarkar2024llmtraining,
  author = {Abhijoy Sarkar},
  title = {LLM Training: A Modular Framework for Fine-tuning Large Language Models},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub Repository},
  howpublished = {\url{https://github.com/acebot712/llm_training}},
}

Contact