Nabugu -

AI groups work to tune, release large language models

CarperAI, Scale AI and other research groups have partnered to deliver instruction-tuned language models as the NLP market grows.

A group of AI researchers have joined forces to train and publicly release instruction-tuned large language models.

CarperAI, a research organization within the EleutherAI research group, has partnered with Scale AI, EleutherAI, Humanloop, Hugging Face and Multi to democratize large language models (LLM). EleutherAI is a research initiative supported by Stability AI, the startup behind Stable Diffusion, a text-to-image AI generative model.

On Wednesday, the organizations revealed an initiative to train and release instruction-tuned LLMs. The models are trained through reinforcement learning from human feedback (RLHF). The open source release will enable academics, researchers and startups to experiment with the models.

The more folks you have doing this as a passion project, the more effective you are because language models generally take a lot of time.
Will McKeon-WhiteAnalyst, Forrester

The new initiative comes as the popularity of language models and natural language processing (NLP) tools and technologies grow. Tuning is one of the biggest issues with NLP tools and technologies, said Will McKeon-White, an analyst at Forrester. Tuning the technology to determine whether a conversation was good or bad requires dedicated staffing within organizations. This can require skills and investment from enterprises.

"Open sourcing is a very effective way to expand the audience of testers," McKeon-White said. "The more folks you have doing this as a passion project, the more effective you are because language models generally take a lot of time."

Having human testers overcome some of the restrictions of large language model tools, such as whether utterances are stated correctly, will save organizations time, he added.

The models will be trained by EleutherAI and Multi, a startup working to apply LLMs to automation. Hugging Face, a community and data science platform, Scale AI and Humanloop will fine-tune the models.

While open source projects are good for expanding the testing group, they also require oversight and administration, McKeon-White said. This will keep bad actors from misusing the models.

Initiatives like these are one telling signs of the interest in NLP technologies, while another is interests from investors. Despite the economic uncertainty, investors continue to place their bets on the NLP market, McKeon-White said.

For example, Stability AI revealed on Monday that it raised $101 million in funding. The vendor launched its free open source text-to-image generator in August. Stable Diffusion has been downloaded by more than 200,000 developers since its launch, according to the vendor.

"We are seeing interest in this space, and organizations are continuing to raise money," McKeon-White said. "It is an emerging area that has very high potential. But the theoretical amount of value that can be garnered is still significantly higher than what is currently being achieved."

Next Steps

Startup’s latest large language model rivals GPT-3

Dig Deeper on AI technologies

Business Analytics
Data Management