Hugging Face, an artificial intelligence (AI) company, offers a range of tools for data research and development, including web dashboards that showcase AI-powered applications, models, and datasets. Despite this wide array of tools, some of the most notable and effective solutions from Hugging Face originate from a small team established in January.
Known as H4 (short for “helpful, honest, harmless, and huggy”), this two-person team is focused on creating tools and “recipes” to enable the AI community to develop AI-driven chatbots akin to ChatGPT. Lewis Tunstall, a machine learning expert at Hugging Face and one of the core members of H4, attributes the inception of this initiative to the inspiration derived from ChatGPT’s advancements.
In a TechCrunch interview, Tunstall mentioned that they commenced planning to replicate ChatGPT’s functionalities using open-source libraries and frameworks following its release by OpenAI in late 2022. H4’s primary research area revolves around instructing Large Language Models (LLMs) on how to respond to human or AI-generated prompts effectively.
H4 has been instrumental in developing various open-source large speech models, including Zephyr-7B, a specialized chat-oriented version of the Mistral 7B model recently introduced by the French AI startup Mitral. Additionally, they enhanced the Falcon-40B, a model from the Technology Innovation Institute in Abu Dhabi, to provide more nuanced responses in natural language interactions.
Similar to other research teams within Hugging Face, H4 leverages a dedicated cluster of over 1,000 Nvidia A100 GPUs for training its models. While physically situated in Europe, Tunstall and his colleague Ed Beeching receive support from internal Hugging Face teams for tasks such as design testing and evaluation.
Beeching emphasized that the deliberate small size of H4 enables them to maintain flexibility and adaptability in a rapidly evolving research landscape. They also collaborate with external organizations like LMSYS and LlamaIndex on shared initiatives.
Recently, H4 has explored various training methodologies and tool development processes to validate their effectiveness in real-world scenarios. They published a comprehensive guide this month containing all the source code and data components used in creating Zephyr. The team plans to update this guide with code from their upcoming AI models as they are released.
When questioned about pressure from Hugging Face’s leadership to commercialize their work, Tunstall clarified that H4 does not directly monetize its tools. However, these tools play a role in Hugging Face’s Expert Acceleration Program, which provides business-focused guidance on developing unique AI solutions.
Beeching stated that H4’s objective is not to compete with other open-source AI projects like EleutherAI and LAION but rather to empower the AI community by sharing training code and data models for their conversational models. He emphasized the importance of community contributions in enabling their work to progress successfully.
