Fine-Tuning

Process models published

Also known as: Fine-tuning, Model Fine-tuning, Transfer Learning

Definition

The process of further training a pre-trained model on a specific dataset to adapt it for a particular task or domain. Fine-tuning updates model weights (unlike prompting, which only changes inputs), enabling deeper specialization. It trades generality for improved performance on the target domain.

What this is NOT

  • Not prompting (fine-tuning changes weights; prompting doesn't)
  • Not pre-training from scratch (fine-tuning starts from pre-trained weights)
  • Not RAG (RAG adds context; fine-tuning modifies the model)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Training a base LLM on custom data to improve performance on specific tasks, adapt to domain terminology, or modify behavior. Done through APIs (OpenAI, Anthropic) or locally with open-source models.

Sources: OpenAI Fine-tuning documentation, Hugging Face PEFT documentation, LoRA and adapter methods

ml-research

A transfer learning technique where a model pre-trained on a large dataset is further trained on a smaller task-specific dataset, transferring learned representations to the new domain.

Sources: Transfer learning literature, BERT fine-tuning paper

Examples

  • Fine-tuning Llama on medical literature for healthcare applications
  • OpenAI fine-tuning API to create a custom assistant
  • LoRA fine-tuning to adapt style with minimal compute
  • Fine-tuning on company documents for internal Q&A

Counterexamples

Things that might seem like Fine-Tuning but are not:

  • Prompting with examples (no weight updates)
  • RAG (adds knowledge without changing model)
  • Pre-training a model from scratch

Relations

  • inTensionWith in-context-learning (Alternative approaches to task adaptation)
  • overlapsWith instruction-tuning (Instruction tuning is fine-tuning for instruction following)
  • overlapsWith rlhf (RLHF is fine-tuning with human preference data)
  • requires foundation-model (Fine-tuning starts from a foundation model)

See Also

Implementations

Tools and frameworks that implement this concept: