Databricks Inc, a big data analysis company, has recently joined the competition in the field of generative AI and released a Dolly The open source large-scale language model of Dolly, named Dolly, is a tribute to Dolly, the first cloned sheep.
Generative AI such as ChatGPT and Bard usually use data from thousands of different websites. The amount of data they use is amazing. In addition, thousands of powerful GPUs are needed to provide support for training AI with these data.
Databricks hopes that through open source Dolly and its training data, anyone can develop an AI that is truly human without investing millions of dollars, and that this kind of AI will no longer be affordable only for large technology companies, and millions of small companies will also be able to benefit from it.
In addition, let each small company build its own model without sending data to the service provider behind the proprietary model through API, which can also protect the sensitive data and proprietary intellectual property of small companies. In addition, some companies may have different trade-offs in model quality, cost and expected behavior, and these companies can also adjust the model according to needs.
Ali Ghodsi, CEO of Databricks, said Dolly only needs very little data and very short time to complete the training.
For only $30, a server and three hours, we can teach Dolly to start human level interaction.
Databricks did not create its own model from scratch, but based on the natural language processing model GPT-J, which EleutherAI will open source in 2021. GPT-J is a natural language processing AI model based on GPT-3 and composed of 6 billion parameters. The model is trained on an 800GB open source text dataset, and can be comparable to the GPT-3 model of similar scale.
Databricks said that although the model is much smaller, with only 6 billion parameters, and smaller data sets and training time (ChatGPT's parameters are 175 billion), Dolly still showed the same "magical human interaction ability" shown by ChatGPT.
Databricks is currently in GitHub A simple Databricks notebook has been opened on Databricks. Developers can use it to build Dolly on Databricks.