Yannic Kulcher Introduces OpenAssistant:
Machine learning researcher and YouTube celebrity Yannic Kilcher announces OpenAssistant which, if it’s as described, is an exciting development in open source generative AI.
OpenAssistant is an open source chat-based assistant in development. The goal is a system that understands tasks, can interact with third-party systems, and can retrieve information from on-line in pursuit of its goals. He also tells us that OpenAssistant will be easily extended and personalized. The projects focusses on putting the human first, being fast and efficient, having a unified direction, being pragmatic, and using consumer-level hardware. Main efforts include collecting data, creating a dataset, making data collection more fun, training models, setting up infrastructure, and ensuring privacy and safety.
Kilcher provides us with this expository graphic outlining OpenAssistant:
OpenAssistant uses InstructGPT’s architecture and workflow as a starting point, including reinforcement learning from human feedback, or RLHF. Candidates for fine-tuning include GPT-J, CodeGen, FlanT5, or GPT-JT.
After joining the project, you can help them in their goal need of collecting human demonstrations of assistant interactions. After you sign in and enter the website, you are faced with this dashboard:
The goal is to collect 50,000 high-quality human-generated Instruction-Fulfillment (prompt and response) samples through a crowdsourced process. The prompts will be reviewed to avoid toxic, spam, junk, and personal information data. For each prompt, multiple completions will be gathered, and users will be asked to rank them. Multiple votes from users will measure overall agreement. The ranking data will train a reward model. In spirit of gamification, there’s even a leaderboard. Top contributors will receive rewards.
The video offers several examples of how to use the tool. Comments on the video include some concerns about guarantees that the project will, in fact, remain democratic: open source and able to run affordably.
