There’s a new player in the language model market trying to challenge OpenAI’s dominance.
On Wednesday, Databricks, a San-Francisco based start-up that develops software used to construct AI systems, publicly released a substantial data set that can be used to train chatbots like ChatGPT — for free. With the open-source release of Dolly 2.0, Databricks has emerged as a fresh contender in the language model industry, aiming to compete with OpenAI’s dominance.
Databricks Unveils Dolly 2.0
In a move company officials said is aimed at making artificial intelligence more accessible for all, Databricks released Dolly 2.0, a large language model (LLM) trained on a human-generated dataset, curated from its own employees, and licensed for research and commercial use.
“We are open-sourcing the entirety of Dolly 2.0, including the training code, the dataset, and the model weights, all suitable for commercial use,” company officials said in a statement. “This means that any organization can create, own, and customize powerful LLMs that can talk to people, without paying for API access or sharing data with third parties.”
According to Databricks, Dolly 2.0 is a language model with 12 billion parameters, built on the EleutherAI pythia model family, that has been exclusively fine-tuned on a new, premium-quality dataset comprising human-generated instructions, crowdsourced from Databricks employees.
Related Article: How to Choose the Best Chatbot for Your Business
Databricks’ New Language Model Dolly 2.0 Aims to Disrupt OpenAI’s Reign
The announcement comes just two weeks after the launch of Dolly, an LLM trained on ChatGPT data, that couldn’t be employed in commercial applications because OpenAI policies prevent the usage of its data to build any commercial AI systems that could rival its offerings.
“When we first released Dolly two weeks ago, we were immediately flooded with requests from people who wanted to try it out. The number one question was —can I use this commercially?” Databricks CEO Ali Ghodsi said. “Dolly was originally trained using a dataset that the Stanford Alpaca team had created with the OpenAI API (and its terms of service seek to prevent anyone from creating a model that competes with OpenAI), the answer was unfortunately, no.”
And that, Ghodsi said, led the company to create their own dataset (databricks-dolly-15k), that could be commercially used, so Dolly 2.0 could be fine-tuned exclusively on a new, human-generated dataset.
“We’re deeply committed to making it simple for customers to use LLMs, so expect both a continued investment in open source, as well as innovations that help accelerate the application of LLMs to key business challenges,” Ghodsi said. “We also hope that the community takes the baton and helps improve our dataset and builds better models that can solve various tasks for businesses and organizations.”
Generative AI for the People
Andy Thurai, VP and principal analyst at Constellation Research, Inc., said be believes the company’s aim in providing open-source data and access to Dolly 2.0, is to help the machine learning community, so they consider using the Databricks platform to create their enterprise-specific, smaller variation of LLMs, instead of depending on large scale LLMs offered by Microsoft or Google.
“While this corpus of this data and the parameters are very small when compared to GPT-3 and Bard models,” Thurai said, “this is meant to demonstrate the creation of localized, individualized versions of LLMs is possible.”
During an interview on the Fortt Knox show, Ghodsi shared his belief that AI and machine learning have the potential to tackle some of the most challenging problems faced by the world today. And through the launch of Dolly 2.0, Ghodsi aims to provide greater public access to technology that is traditionally owned and controlled by private entities.
“Most of these models today have either been proprietary — so very few companies have had them — or there are some open-source ones, but they’re not good. You can’t use them for this kind of human level interaction,” Ghodsi said. “Today, we’re releasing for the first time a model that can do this kind of human interaction, it’s called instruction following, and its completely open source and you can use it commercially … we’re excited about that … we’re excited what organizations out there can do with this and what problems that can solve.”
Have a tip to share with our editorial team? Drop us a line: