Dolly 2.0, like its predecessor Dolly released a few weeks ago, uses a smaller dataset than most major language models. Dolly had 6 billion parameters, and in Dolly 2.0 they are twice as many – 12 billion. For comparison, GPT-3 has 175 billion parameters. Dolly 2.0 is reportedly built on a high-quality dataset.

What the new language model offers

A distinctive feature of the new generative AI models is the ability to use their own training datasets to generate coherent sentences and answers to user questions. And Dolly 2.0 can do this even with a much smaller output than OpenAI models. This allows you to use the model on your own servers without having to share data with third parties.

We believe models like Dolly will help democratize LLM by transforming it from something that not all companies can afford, to something every company can own and customize to improve their products.
– Indicated in Databricks.

In a comment to SiliconANGLE, the head of Databricks emphasized: businesses can “make money from Dolly 2.0.” The company offers Dolly 2.0 under a fully open-source Creative Commons license and the databricks-dolly-15k training dataset containing 15,000 high-quality man-made query-response pairs. All of these can be freely used, modified and added, as well as used in commercial projects at no cost to anyone.

According to Databricks, Dolly 2.0 is currently the only model without license restrictions. Other models such as Alpaca, Koala, GPT4All and Vicuna cannot be used for commercial purposes due to the use of training data provided to them with certain conditions.

Interestingly, the first version of Dolly was trained on Stanford Alpaca data using the OpenAI API, so it was not commercially available.