A framework streamlining Training, Finetuning, Evaluation and Deployment of Multi Modal Language models
- Diverse Model Support: Llama3, Phi, Mistral, Gemma, and more.
- Versatile Image Encoding: CLIP, Seglip, RADIO, and others.
- Customization Made Simple: YAML files and CLI for adaptability.
- Efficient Resource Utilization: Seamless operation on a single GPU.
- Seamless Deployment: Docker locally or on cloud with Skypilot.
- Comprehensive Documentation: Includes datasets for successful implementation.
- Introduction
- Supported_Models
- Changelog
- Installation
- Pretrain
- Finetune
- Evaluate
- Inference
- Features to be Added
- Citation
- Acknowledgement
- Llama3
- Phi
- Mistral
- Gemma
- Version 1.0.1:
- Added support for distributed training.
- Included accelerate library.
- Version 1.0.0:
- Initial release.
- Clone the repository from GitHub.
- Install dependencies using pip:
pip install -r requirements.txt
. - Run
setup.sh
to set up the environment. - Start using Eagle!
- Utilize supported models for pretraining multimodal models.
- Fine-tune pretrained models with custom datasets or tasks.
- Evaluate model performance using specified metrics and datasets.
- Deploy models for inference on new data or integrate them into existing systems.
- Add support for accelerate.
- Add support for additional Huggingface models such as falcon, mpt.
@article{AdithyaSKolavi2024,
title={Eagle: Unified Platform to train multimodal models},
author={Adithya S Kolavi},
year={2024},
url={https://github.com/adithya-s-k/eagle}
}
We would like to express our gratitude to the creators of LLaVA (Large Language and Vision Assistant) for providing the groundwork for our project. Visit their repository here.