Fine-Tuning

Definition

Training a pre-trained AI model on specialized data to improve performance on specific tasks.

Why It Matters

Base models are generalists. When you need a model that consistently writes in your brand voice, classifies your specific 50 ticket categories, or knows your internal product manual, fine-tuning is the bigger lever than prompting alone. The cost: an upfront training run and a custom checkpoint to host.

Key Points

• QLoRA (Quantised Low-Rank Adaptation) makes fine-tuning a 7B model feasible on a single 24 GB GPU, the method most open-source fine-tuners use today.
• LoRA adds trainable rank-decomposition matrices on top of frozen base weights, typically at rank 8–64; only those matrices are updated during training.
• Rule of thumb: 1K high-quality examples is enough for style/format fine-tuning; 10K–100K is needed for capability extension.
• RLHF (Reinforcement Learning from Human Feedback) is fine-tuning where the reward signal comes from human preference ratings, how GPT-4 and Claude are aligned.
• Full fine-tuning a 70B model at FP16 needs ~560 GB of VRAM for weights plus optimizer states, only tractable on a multi-GPU cluster.

Example

Take a 7B open-source model, train it for 3 epochs on 10K customer-support transcripts paired with their resolutions, and you get a model that drafts plausible first-response replies for that specific company at zero inference markup.

Related Terms

• LLM (Large Language Model)A neural network trained on massive text datasets that can generate, understand and manipulate human language. Examples: GPT-4, Qwen, Claude.
• ParameterA trainable weight in an AI model. Larger models have more parameters (7B, 70B, 400B).
• Open Source AIAI models released with open licenses (MIT, Apache 2.0) allowing anyone to use, modify and deploy them.

Quick Facts