GitHub - IBM/Dromedary: Dromedary: towards helpful, ethical and reliable LLMs.

🐪 Explore Dromedary from IBM - an open-source, self-aligned language model trained for helpful, ethical, and reliable Large Language Models. Dromedary-2 introduces new features for enhanced performance and self-alignment process! #AI #LLM #EthicalAI

Dromedary is an open-source, self-aligned language model trained with minimal human supervision.
Dromedary-2 introduces a new self-alignment process involving diverse user prompts and exemplars to enhance performance without verbose cloning or inference-time few-shot examples.
Dromedary-2 also offers the RLAIF training pipeline for self-alignment with principle-following reward models.
Training your own self-aligned model or performing inference with quantities differing from 1, 2, 4, or 8 can be done using the llama_dromedary package.
The Dromedary weights are released as delta weights, which can be added to the original LLaMA weights.
Synthetic data for self-alignment training is released in Hugging Face Datasets Hub.
The project provides a chatbot demo, full training pipeline, and human annotations.
The paper for Dromedary should be cited when using the data or code.
Acknowledgements are extended to various open-source efforts in the domain of large language models like Meta LLaMA team, Standford Alpaca team, and Hugging Face PEFT.
Dromedary aims towards the development of helpful, ethical, and reliable Large Language Models (LLMs).