AI startup Hugging Face delivers an extensive array of data science hosting and development tools, featuring a GitHub-like portal tailored for AI code repositories, models, and datasets. Additionally, they provide web dashboards for demonstrating AI-powered applications. Notably, some of Hugging Face’s most remarkable and potent tools have emerged from a two-person team formed as recently as January.
This team, known as H4, short for “helpful, honest, harmless, and huggy,” is dedicated to developing tools and “recipes” to empower the AI community in creating AI-driven chatbots akin to ChatGPT. The inception of H4 was prompted by the release of ChatGPT in late 2022. Lewis Tunstall, a machine learning engineer at Hugging Face and one of H4’s members, explained that their primary research focus revolves around alignment, aiming to teach Large Language Models (LLMs) how to behave based on feedback from humans or other AIs.
H4 has played a pivotal role in the creation of various open source large language models, including Zephyr-7B-α, a chat-centric version of Mistral 7B from Mistral, and a modified Falcon-40B from the Technology Innovation Institute in Abu Dhabi, adapted to respond more helpfully to natural language requests.
The team, consisting of Tunstall and Ed Beeching, operates remotely in Europe and relies on support from internal Hugging Face teams, such as model testing and evaluation. Despite its small size, H4 deliberately stays agile to adapt to the evolving research landscape and collaborates externally with groups like LMSYS and LlamaIndex.
H4’s recent endeavors involve exploring alignment techniques and constructing tools to assess the efficacy of techniques proposed by the broader community and industry. They have released a handbook containing source code and datasets for Zephyr, with plans to update it as they release future AI models.
Regarding commercialization pressure from Hugging Face’s higher-ups, Tunstall clarified that H4 doesn’t directly monetize its tools. However, the tools contribute to Hugging Face’s Expert Acceleration Program, an enterprise-focused offering providing guidance for building custom AI solutions.
When asked about competition with other open source AI initiatives, Ed Beeching emphasized that H4’s objective is not competition but empowerment. Their aim is to empower the open AI community by releasing training code and datasets associated with their chat models, acknowledging the collaborative nature of their work with contributions from the community.
