Databricks developed the open-source code Dolly for two years, acquiring 6 billion parameters from EleutherAI and using data from Alpaca to bring ChatGPT-like interactivity.
Heaptalk, Jakarta — A startup headquartered in San Francisco, Databricks, introduced an open-source code to build an AI chatbot, namely Dolly. Claimed to be a cheap-to-build large language model (LLM), Dolly shows off instruction-following capabilities similar to ChatGPT developed by OpenAI.
The AI model acquires 6 billion parameters, far less than GPT-3 which has 175 billion parameters. However, Dolly, which has only been developed for two years, is considered to work very well. However, Reuters reported that Databricks has yet to release a formal benchmark test to show Dolly’s performance equal to ChatGPT.
Dolly harnesses an existing open-source parameter model from EleutherAI and adjusts it by degrees to bring out instruction-following capabilities, including brainstorming and text generation not present in the original model, utilizing data from Alpaca. Alpaca is an instruction-following model developed by the Center for Research on Foundation Models (CRFM) at Stanford University.
“We are calling the model Dolly—after Dolly the sheep, the first cloned mammal—because it’s an open-source clone of an Alpaca, inspired by a LLaMA. We are in the earliest days of the democratization of AI for the enterprise, and much work remains to be done, but we believe the technology underlying Dolly represents an exciting new opportunity for companies that want to cheaply build their own instruction-following models,” stated Databricks on its official blog page.
Quoted by Reuters, CEO of Databricks Ali Ghodsi conveyed that the release was aimed at demonstrating a viable alternative to training a kind of AI model called a large language model with enormous resources and computing power. OpenAI charges access to its models for enterprise-owned applications. By 2024, the ChatGPT developer is targeting $1 billion in sales.
By launching this open-source code, Databricks, which was founded in 2013, wants companies to train their own AI models using its software. “The future will be that everyone has their own model, and they can actually train it, and they can make it better. And that way, they also don’t have to give away their data to someone else,” concluded Ali.
