Full Fine-Tuning Update all parameters. Expensive but most powerful. LoRA (Low-Rank Adaptation) Train only small adapter matrices. 10-100× cheaper. QLoRA LoRA + 4-bit quantization. Runs on consumer GPUs. DPO (Direct Preference Optimization) Replace RLHF with simpler preference learning.

OpenAI - Models: GPT-4o-mini, GPT-3.5 - Cost: $25/M training tokens - Pros: easiest, hosted - Cons: limited models, closed Anthropic - Limited fine-tuning (RLHF mostly) - Best for: alignment work - Cons: less ecosystem Cohere - Multilingual fine-tuning - Pros: Turkish strong - Cost: similar to OpenAI Together / Fireworks / Anyscale - Open-source models fine-tuned hosted - More flexibility, slightly more work Self-Hosted - Free + you control everything - Need: Axolotl, Unsloth, or vLLM - Hardware: 1× A100 minimum for 7B

Quality Quantity 1000 high-quality 10K mediocre. Format Most providers want JSONL: jsonl {"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]} Diversity Spread across edge cases. If 80% of training is happy path, model struggles at edges. Eval Split Reserve 10-20% as eval set. Never train on eval.

Hyperparameter Tuning

Learning rate: start 1e-5 (OpenAI), 1e-4 (LoRA) - Epochs: 3-5 default. More = overfit risk - Batch size: max your hardware allows - LoRA rank: 8-16 typical

OpenAI Fine-Tuning Dashboard: easiest - Together: open-source hosted - Axolotl: most flexible YAML config - Unsloth: 2× faster than Axolotl - HuggingFace AutoTrain: GUI option

Build eval set BEFORE fine-tuning - Score base model on it - Score fine-tuned on same set - Measure: accuracy, latency, cost, qualitative

Overfitting: dataset too small, model memorizes - Catastrophic forgetting: model loses general ability - Bias amplification: training data bias gets stronger - Style mismatch: brand voice drifts during training

For 50K training tokens: - OpenAI GPT-4o-mini: $1.25 - Together Llama 8B: $2 - Self-hosted A100 (1 hour): $2-5 cloud rental Total to fine-tune: $5-50 typical.

Fine-Tuning Complete Guide 2026 — Guide

Fine-Tuning Complete Guide 2026

When to fine-tune, how to do it, what avoids burning money. When to Fine-Tune - Style consistency: brand voice, tone - Structured output: strict JSON / format - Domain language: medical, legal jargon - Cost optimization: smaller fine-tuned beats general bigger - Latency: fine-tuned smaller = faster When NOT to Fine-Tune - Dynamic knowledge: use RAG - Less than 100 examples: not enough signal - Ju…

All guides

Related guides

AI for Edge Computing 2026
AI Inference Optimization 2026
AI Prompt Engineer 2026
Astro 5 Complete 2026
Cursor vs Copilot 2026