Dolly v2 12B

Dolly 2.0 is Databricks' 12 billion parameter instruction following model, released in April 2023 and notable as the first instruction tuned LLM with full commercial viability. It builds on EleutherAI's Pythia-12B and is fine tuned on databricks-dolly-15k, a dataset of 15,000 human written instruction response pairs crowdsourced from Databricks employees.

Because earlier efforts like Alpaca depended on OpenAI generated data with restrictive terms, Dolly 2.0's fully human, openly licensed pipeline was a breakthrough: model, code, dataset, and weights all released under permissive terms including MIT for the model. It showed organizations could own and customize capable chat models without API dependence.