A new paper titled “H2O Open Ecosystem for State-of-the-art Large Language Models” introduces two open-source libraries for easily developing and evaluating large language models (LLMs). The goal is to offer customizable alternatives to proprietary systems like GPT-3 and ChatGPT.
The researchers at H2O.ai release h2oGPT (https://github.com/h2oai/h2ogpt) – a library for benchmarking and comparing LLMs, and H2O LLM Studio (https://github.com/h2oai/h2o-llmstudio) – a graphical user interface for efficient LLM fine-tuning without coding. Together these no-code tools allow anyone to optimize, train and deploy customized LLMs locally.
H2O LLM Studio supports state-of-the-art techniques like LoRA adapters and reinforcement learning. It enables fine-tuning popular open LLMs up to 70 billion parameters through an intuitive GUI. h2oGPT allows evaluating multiple models concurrently.
According to the authors, easily tailored open-source LLMs can enhance transparency, innovation and responsible AI development. The no-code tools give full control over privacy, behavior and carbon footprint. Democratized access helps accelerate applications in chatbots, summarization, code generation and more.
But critical challenges remain around harmful model biases and content. Continued open research is key to steering LLMs to benefit society responsibly.