Dromedary is an open-source large language model project from IBM Research, aimed at building helpful, ethical and reliable LLMs with a minimal amount of human supervision. The GitHub repository, which has roughly 1,100 stars, releases the models, code and methodology behind the Dromedary family.
Dromedary explores how far alignment techniques can go with very little human labelling. Its training recipe combines principles-guided self-alignment โ where the model learns to follow a short set of written principles โ with limited human annotations, producing instruction-following models based on open foundations such as LLaMA. The project published model weights and detailed documentation of the approach, making it a reference point in the early open-LLM alignment literature. The repository explains the training stages, provides inference code, and links the associated research, so others can reproduce or build on the self-alignment pipeline. While the ecosystem has since moved to much larger models, Dromedary remains a compact, well-documented case study in doing alignment cheaply and transparently.
Dromedary’s releases include variants trained with different amounts of human annotation, letting researchers study the effect of supervision on alignment quality directly. The project is frequently cited in surveys of open instruction-tuned models from the first generation of LLaMA-based assistants.
Dromedary is free and open source under the GPL-3.0 license; the models and code can be downloaded from GitHub at no cost, subject to the terms of the underlying base models as well.
AI researchers and practitioners interested in LLM alignment techniques, especially those studying how self-alignment and small amounts of supervision can produce well-behaved open models.