📚 Glossary

Fine-Tuning

Fine-tuning is the process of further training a pre-trained language model on a smaller, task-specific dataset to improve its performance on a particular type of task or domain.

Definition

Fine-tuning is a machine learning technique where a model that has already been trained on a large general corpus (the base model) is further trained on a smaller, curated dataset focused on a specific task, domain, or style. The additional training adjusts the model's weights to improve performance on the target task without starting from scratch.

Pre-training vs. Fine-tuning

Pre-training is the initial, large-scale training phase where the model learns general language understanding from massive text corpora. This is computationally expensive and done once.

Fine-tuning happens after pre-training and refines the model for specific applications. It's faster and cheaper because the model already understands language — fine-tuning just adjusts its behavior for a narrower task.

Examples of Fine-Tuning

  • Training a base LLM on medical literature to improve performance on clinical documentation tasks
  • Training on customer service conversations to create a domain-specific support chatbot
  • Training on legal documents to improve contract analysis accuracy
  • Training on examples of humanized text (AI input → human output pairs) to improve an AI humanizer's output quality

Fine-Tuning vs. Prompt Engineering

Both techniques improve model output quality, but they work differently:

  • Prompt engineering shapes the model's behavior through carefully crafted instructions at inference time — no training required
  • Fine-tuning shapes the model's behavior by updating its weights — requires a training dataset and compute

Fine-tuning tends to produce more consistent results for specific tasks but requires more resources. Prompt engineering is more flexible and faster to iterate.

Instruction Fine-Tuning and RLHF

Modern chat models like ChatGPT are not just pre-trained but also fine-tuned through instruction fine-tuning (training on example instruction-following) and RLHF (Reinforcement Learning from Human Feedback), where human raters score model outputs and those scores are used to further train the model to produce preferred outputs.

Large Language Model · Training Data · Prompt Engineering