🎛️ Fine-Tuning & Custom Models
Learn Fine-Tuning & Custom Models on AI4AI — short, hands-on lessons with live AI runs, at three reading levels (beginner to expert). Free to start.
What you'll learn
- Prompting vs RAG vs Fine-Tuning: Pick the Right Tool (7 min) — When you need an AI model to behave differently, you have three tools, each at a different cost and complexity level. *…
- When Fine-Tuning Helps — and When It Doesn't (7 min) — Fine-tuning retrains a model's weights on a curated dataset to shift its default behavior. It works well for three thin…
- Building a Fine-Tuning Dataset (7 min) — A fine-tuning dataset is a curated collection of input-output pairs that teach a model a specific behavior. Each exampl…
- Running a Fine-Tune: Model Choice, Hyperparameters, and Overfitting (8 min) — When running a fine-tune, three decisions dominate outcomes. **Base model choice** sets your ceiling. Instruction-tuned…
- Did Fine-Tuning Actually Help? Evaluating Your Custom Model (8 min) — Evaluating a fine-tuned model requires three parallel assessments: task performance, general capability retention, and …
- Before You Fine-Tune: Prompt Optimization, LoRA, and Distillation (8 min) — Full fine-tuning is expensive: GPU hours, curated datasets, and ongoing maintenance. Before committing to it, three che…
Lessons
Prompting vs RAG vs Fine-Tuning: Pick the Right Tool
When you need an AI model to behave differently, you have three tools, each at a different cost and complexity level. **Prompting** means adding instructions, context, or examples directly to your input — the system prompt or user message. It costs nothing extra, takes seconds t…
When Fine-Tuning Helps — and When It Doesn't
Fine-tuning retrains a model's weights on a curated dataset to shift its default behavior. It works well for three things: consistent tone and style (always formal, always concise), structured output format (valid JSON matching a fixed schema, specific templates), and narrow rep…
Building a Fine-Tuning Dataset
A fine-tuning dataset is a curated collection of input-output pairs that teach a model a specific behavior. Each example is a demonstration: given this input, produce exactly this output. The model learns to replicate the pattern across all examples, so the patterns must be cohe…
Running a Fine-Tune: Model Choice, Hyperparameters, and Overfitting
When running a fine-tune, three decisions dominate outcomes. **Base model choice** sets your ceiling. Instruction-tuned checkpoints—models like Llama-3.1-8B-Instruct or Mistral-7B-Instruct—are the practical starting point for most tasks. They already follow directions, so you're…
Did Fine-Tuning Actually Help? Evaluating Your Custom Model
Evaluating a fine-tuned model requires three parallel assessments: task performance, general capability retention, and cost-efficiency. Measuring accuracy on your target task alone is not enough — you must also run both models against general benchmarks (e.g., MMLU, TruthfulQA) …
Before You Fine-Tune: Prompt Optimization, LoRA, and Distillation
Full fine-tuning is expensive: GPU hours, curated datasets, and ongoing maintenance. Before committing to it, three cheaper interventions should be tried in order. **Prompt optimization** is free and fast. Systematic prompt engineering — detailed system prompts, few-shot example…