InstructLab: Applying Open Source Methods to Building and Trai...
summary
InstructLab claims a workflow to add domain contributions to an existing LLM without full forking or end-to-end fine-tuning, using a pipeline of taxonomy + data generation + validation + multi-stage tuning. It uses separate teacher/critic/student models, generates hundreds–thousands of synthetic QA pairs, and validates for bias and quality before tuning. Tools include llama.cpp, quantized 7B Q4 models on Mac M1/M2 (32 GB), and backend tuning with MLX/Kora, DeepSpeed/Lora; initial full runs used ~128 GPUs and ~25+ hours.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:07AI and LLMs overview
- 2:34LLM tuning challenges
- 8:00The problem of contributing to LLMs
- 10:30Introducing LAB technique
- 15:27LAB skills and knowledge
- 21:06Additive model updates with LAB
- 22:12InstructLab community
- 25:45InstructLab workflows and testing
- 31:48Live demo and CLI
- 36:29Summary and future
related talks
40mInstructLab: Applying Open Source Methods to Building and Trai...
28mA community based open source LLM and what it can do for your business
30mApplying Open Source Methods to Building and Training Large Langu...
29mLocal LLM fine-tuning: a practical example
01h 49mAI engineering with open access LLMs that lie, curse, and steal
15mOpenAssistant's Opensource LLMs Saga
59mWorkshop: Choosing the Best Open Source LLM for Your Application
48mPracticalities of Fine-Tuning Llama 2 with AI Studio