When should I fine-tune vs use prompt engineering?

Fine-tune when: you have at least 500+ high-quality examples, need consistent output formatting, want to reduce prompt length/cost, or need the model to learn domain-specific knowledge. Use prompt engineering for most cases - it's cheaper and faster to iterate. Only fine-tune when prompt engineering hits a ceiling.

How much training data do I need for fine-tuning?

Minimum: 100-500 examples for simple tasks. Ideal: 1,000-10,000 examples for reliable performance. OpenAI recommends at least 50-100 well-curated examples to see meaningful improvement. Quality > quantity - 500 excellent examples outperform 5,000 noisy ones.

What factors affect fine-tuning cost most?

Dataset size (more tokens = higher cost), number of epochs (more epochs = higher cost), and model size (GPT-4o costs 3x more than GPT-3.5 to fine-tune). Inference cost depends on monthly usage volume and whether you self-host or use API.

Is fine-tuning worth the cost vs using a base model?

For most applications, no - prompt engineering with a good base model (GPT-4o, Claude Sonnet) achieves 90%+ of fine-tuning results. Fine-tuning becomes worth it when you need: consistent output format, reduced latency (shorter prompts), domain-specific knowledge, or better performance on specialized tasks.

Can I fine-tune open-source models to avoid per-token costs?

Yes. Fine-tuning Llama 3, Mistral, or DeepSeek models costs GPU compute ($1-5/hour) and then you own the model weights. Inference has near-zero per-token cost (just GPU compute). Break-even vs API-based fine-tuning is typically 1-5M inference tokens/month. Use [Ollama](https://ollama.ai) or [vLLM](https://vllm.readthedocs.io) for self-hosting.

How does LoRA reduce fine-tuning costs?

LoRA (Low-Rank Adaptation) fine-tunes a small number of parameters instead of all of them, reducing training cost by 80-90% and storage by 99%. LoRA adapters can be swapped at inference time without multiple model deployments. Most modern fine-tuning frameworks use LoRA or QLoRA (quantized LoRA).

What is the difference between fine-tuning and RAG?

Fine-tuning changes the model's weights to embed domain knowledge. RAG (Retrieval-Augmented Generation) keeps the model unchanged and provides relevant documents in the prompt. RAG is cheaper, more flexible, and easier to update. Fine-tuning is better when: latency matters, you need the model to internalize patterns, or prompts are too long.

How do I evaluate fine-tuning ROI?

Compare performance (accuracy, relevance, format compliance) before and after fine-tuning on a held-out test set. Calculate cost savings from shorter prompts (fine-tuned models need fewer few-shot examples). Measure improved user satisfaction or task completion rates. If fine-tuning improves accuracy by 20%+ or reduces prompt costs by 50%+, it's likely worth it.

AIファインチューニング費用計算機

AIモデルのファインチューニング費用をトレーニング費用と推論費用を含めて計算します。