dolly

Dolly is an instruction-following language model released by Databricks in 2023. It is built on top of an existing model called Pythia-12B and trained further on a dataset of roughly 15,000 question-and-answer pairs written by Databricks employees. The result is a model that can follow written instructions in a way the base model could not, and it is licensed for commercial use, which was notable at the time of its release.

The training dataset, called databricks-dolly-15k, covers several categories of tasks: answering questions from a given passage, open-ended question answering, summarization, classification, text generation, information extraction, and brainstorming. The data was written by people, not generated by another AI, and is released under a permissive Creative Commons license.

The repository documents the model's known limitations honestly. Dolly v2 (the 12B parameter version) struggles with complex reasoning, mathematical problems, factual accuracy, and generating responses in specific formats. The maintainers describe it as exhibiting surprisingly good instruction-following behavior for a model of its size, but they are clear that it is not competitive with larger or more recent models.

To run the model yourself, you can load it from Hugging Face using the standard transformers Python library. Full inference requires a GPU with at least 24 GB of memory, such as an A100 or A10. For smaller GPUs, the model can be loaded in 8-bit mode with some quality tradeoff.

Training the model from scratch is also documented for users with access to Databricks and a multi-GPU cluster. The training notebook runs on Databricks notebooks and supports A100, A10, and V100 GPU types, with configuration files provided for each.