OpenAssistant's Opensource LLMs Saga
summary
Claims building top open source LLMs using a pipeline of synthetic dataset creation, fine-tuning and reward modeling then RLHF, producing chat-capable models that follow instructions. Datasets run into hundreds of thousands of instruction–response pairs created cheaply by other LLMs, models scale from about 3B to 70B parameters, and infrastructure scales accordingly. Alignment reduces toxic outputs but sometimes degrades niche skills like code generation. Also producing an open framework called ragas for evaluation and continual improvement that is already in use.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:08OpenAssistant AI vision
- 3:39Building open-source LLMs
- 7:10Fine-tuning for instruction
- 9:42Model alignment
- 12:10Key challenges
- 14:19Future of open-source AI
related talks
01h 49mAI engineering with open access LLMs that lie, curse, and steal
42mOpen Source LLMs — Powering AI Accessibility and Innovation
35mOpen-source AI: why it matters and how to get started
17mArchitecting LLM-Based Chatbots using free open-source LLMs
59mWorkshop: Choosing the Best Open Source LLM for Your Application
40mInstructLab: Applying Open Source Methods to Building and Trai...
29mThe Evolution of LLMs: From Imitation to Reasoning
31mThe AI Revolution Will Not Be Monopolized: How open-source beats economies of scale, even for LLMs