🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
summary
Shows that consumer GPUs plus careful data work let teams skip expensive pretraining and fine-tune open models using synthetic and human-labeled mixes, with concrete numbers: 6–10k high-quality examples for instruction tuning and DPO often working from ~3k. Demonstrates practical tools and failures — Hugging Face training code, GPT-4 for curation, and a classifier trained with ~8 samples per label to clean data — and exposes a bug in UltraFeedback ratings that flipped labels and leaked training data into benchmarks, which gets fixed before training Notus/Notux.
AI-written summary · about
chapters (10)
- 0:00Introduction to LLM training
- 0:50Tools and synthetic data
- 3:32Training phases of LLMs
- 7:53Case study: Data quality
- 14:07Improving annotation quality
- 17:13Human-synthetic data hybrid
- 18:34Advanced data pipelines
- 21:09Best practices for training
- 22:53Additional tools and projects
- 24:03Q&A and concluding remarks
related talks
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
21mBeat GPT 4 and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning
31mReliable Data for Large ML Models: Principles and Practices
42mPost-Training Small Language Models: the adventures of a practitioner
39mCollect Data to Fine-Tune Your Large Language Model
10mFrom Pretrained To Personal: Privacy-First...
33mBuild Better AI with First-Party Data and Fine-Tuning
55mNo Data? No Problem: Zero-Data Model Training Foundational Models