Fine-Tuning LLMs
← Back to Large Language Models
Adapting a pretrained LLM to specific tasks or domains by further training on task-specific data. Ranges from full fine-tuning to parameter-efficient methods.
Methods
- Full Fine-Tuning — update all weights, expensive but most capable
- LoRA (Low-Rank Adaptation) — add small trainable matrices, freeze original weights
- QLoRA — LoRA on quantized models, enables fine-tuning on consumer GPUs
- Adapter Methods — insert small trainable modules between frozen layers
Related
- Foundation Models (starting point for fine-tuning)
- RLHF and Alignment (fine-tuning for alignment)