An open-source language model has emerged that corporations can use for business purposes.
Big data company Databricks has released ‘Dolly 2.0’, the successor to ‘Dolly’, a chatGPT-class enterprise language model that was released two weeks ago, Enterprise Beat reported on the twelfth (local time).
Reportedly, Dolly 2.0 is an open-source language model fine-tuned for a freely available dataset, available for business use. Users can use it in their very own business applications without paying for API access or sharing data with third parties.
Dolly, which was released earlier, is a small language model (sLLM) built by fine-tuning for 3 hours using 50,000 datasets based on ‘GPT-J’, an open source language model with 6 billion parameters developed by Eleuther AI. .
Even smaller than ChatGPT’s 175 billion parameters, the model has ChatGPT-class human interaction capabilities, despite a smaller dataset and training time.
Nonetheless, it couldn’t be used commercially since the dataset used for fine-tuning was not completely open source.
However, Dolly 2.0 is predicated on the ‘Pythia’ open-source language model of Eleuther AI’s 12 billion parameters, and fine-tunes the open-source dataset ‘databricks-dolly-15k’. It’s a coordinated and built sLLM.
Databricks-dolly-15k is a dataset of 15,000 records created by 1000’s of Databricks employees and could be used, modified and prolonged for any purpose, including academic or business applications.
Subsequently, Dolly 2.0 discloses each the model and training data as open sources, allowing corporations to make use of them commercially without limit.
In keeping with Databricks, Dolly 2.0 is currently the one model without this limitation. In actual fact, sLLMs corresponding to Alpaca, Koala, GPT4All and Vicuna are usually not commercially available due to their training data.
“Unlike previous models, Dolly 2.0 comes with a license allowing for business use,” said Databricks CEO Ali Goji. There isn’t a need to offer these datasets to third-party vendors, and models could be monetized in ways traditional models are forbidden. In fact, we wish people to coach models on Databricks, but these models and datasets could be used elsewhere.”
Chan Park, cpark@aitimes.com