Super Data Science: ML & AI Podcast with Jon Krohn 674: Parameter-Efficient Fine-Tuning of LLMs using LoRA (Low-Rank Adaptation)
4 snips
Apr 28, 2023 Learn about Parameter-Efficient Fine-Tuning with LoRA, Atta-Lora, and optimization techniques for large language models. Discover how to reduce trainable parameters and memory usage while adapting fine-tuning in specific model sections for efficiency.
AI Snips
Chapters
Transcript
Episode notes
PEFT Fixes Cost And Forgetting
- Fine-tuning entire LLMs on small data is expensive and causes catastrophic forgetting.
- Parameter-efficient fine-tuning (PEFT) fixes cost and preserves original capabilities while improving small-data performance.
Use Hugging Face PEFT Library
- Use Hugging Face's PEFT library to implement parameter-efficient methods without deep linear algebra expertise.
- These methods store checkpoints in megabytes and avoid retraining full model weights.
LoRA's Low-Rank Trick
- LoRA inserts low-rank decomposition matrices into transformer layers while freezing original weights.
- This reduces trainable parameters ~10,000x and memory footprint about 3x, sometimes improving performance.