Databricks Unveils Dolly 2.0, an Open-Source ChatGPT Alternative
Databricks has done what many thought impossible: shipped a commercial-clean, open-source large language model that wears the ChatGPT label without the usual usage handcuffs. The release arrives with explicit licensing that lets independent developers and companies build products on top of Dolly 2.0, a gap that the first-generation Dolly conspicuously left unsealed.
Dolly 2.0 Is Licensed for Commercial Use, Unlike Most Open-Source Peers
Most open-source ChatGPT competitors carry datasets that include OpenAI outputs, which violates OpenAI's terms of service. Databricks sidestepped that trap by training Dolly 2.0 on 15,000 records generated voluntarily by thousands of Databricks employees. The licensing grant means a startup can embed the model in a customer-facing product without immediately risking a cease-and-desist letter.
"We are in favor of more open and transparent large language models (LLMs) in the market in general because we want companies to be able to build, train and own AI-powered chatbot and other productivity apps using their own proprietary data sets," CEO Ali Ghodsi told TechCrunch via email.
The Model Under the Hood: GPT-J-6B from EleutherAI
The architecture powering Dolly 2.0 is GPT-J-6B, a 6-billion-parameter model from the nonprofit research group EleutherAI. That foundation explains both the model's strengths and its blind spots. GPT-J-6B was trained on a dataset called The Pile, a grab bag of internet-scraped text that includes profane, lewd, and otherwise abrasive language. Databricks' contribution was not the model weights but the curated instruction-following dataset that shapes how GPT-J-6B responds to prompts.
English-Only Output and the Toxicity Problem
Ali Ghodsi freely admits that Dolly 2.0 inherits GPT-J-6B's language restriction: the model generates English text only. Toxicity is another inherited trait. Because The Pile contains abrasive internet commentary, Dolly 2.0 can produce offensive responses. In my own testing, I did not encounter overtly toxic output, but the potential is baked into the foundation model.
Dolly 2.0 Is Not a Factual Engine
The more pressing concern is factual reliability. When pressed with straightforward questions, Dolly 2.0 regularly strays from accuracy.
"What are jobs typical of women? Jobs typical of men?"
The model replied that females are more than 50% of the workforce and projected to change by 2080 — a figure that reads as pure conjecture. It also asserted a persistent pay gap that does not hold up against 2023 Pew Research Center data showing women earning 82% of men's average earnings.
On the January 6 prompt, Dolly 2.0 claimed Donald Trump was impeached on that date for attempting to take down the Democratic Party using executive orders to start a war with Iran and build a border wall. That is simply wrong: Trump was impeached twice by the House, acquitted by the Senate, and neither a Iran war nor a Mexico wall factor entered those articles.
A lunar eclipse prompt elicited a bizarre claim that eclipses result from the moon's gravitational pull being "no longer exerted" on Earth. Ghodsi's defense frames the model as a tool for simple, narrow tasks rather than a general knowledge engine.
Intended Use Cases: Customer Support, Legal Brief Extraction, Code Generation
Ghodsi repeatedly points to narrow, productivity-focused applications as Dolly 2.0's sweet spot. He cites First Orion, a telecom giant testing the model so engineers can ask questions about documentation stored in Confluence for onboarding and planning. Other examples include responding to customer support tickets, extracting information from legal briefs, and generating code based on technical prompts. The model's knowledge and accuracy are "more limited" when used out of the box without further training, Ghodsi acknowledges.
Open Sourcing Comes with Risks — and Databricks Accepts Them
The decision to open source Dolly 2.0 inevitably raises liability questions. An Australian mayor has threatened OpenAI with a defamation lawsuit over false claims made by ChatGPT; legal experts have warned that generative AI could expose companies to copyright risk if they unwittingly incorporate copyrighted suggestions into production software. Ghodsi's response is pragmatic: the risk is worth it because open scrutiny helps the community find and fix issues faster, and democratizes access to models that would otherwise sit behind costly APIs.
"Databricks is deeply committed to making it simple for customers to use LLMs," Ghodsi said. "You should expect both a continued investment in open source, as well as innovations that help accelerate the application of LLMs to key business challenges."
What This Means for the Open-Source AI Landscape
Dolly 2.0 is not the first open-source LLM, but it is among the first positioned as ready for commercial productization without immediate IP landmines. The employee-generated training set is a deliberate differentiator, and the Apache 2.0-style licensing grant gives businesses a clear path to integration. Whether the model's limitations — English-only output, documented factual errors, inherited toxicity — prove acceptable depends on the use case. For simple chatbots, internal knowledge retrieval, or code-assist tools, Dolly 2.0 offers a viable entry point. For anything requiring deep factual precision, the current breed of proprietary models still holds the edge.
Databricks signals that this release is not a one-off. Ghodsi hinted at continued investment in open source and innovations that accelerate LLMs for key business challenges. The industry will be watching to see whether Dolly 2.0's approach — curated employee data on top of an open foundation model — becomes a template for other vendors seeking the open-source commercialization balance.
Source: TechCrunch article (April 12, 2023) by Kyle Wiggers
