Data Requirements & Cost Landscape

How much data you actually need

Quality level	Examples	What you can expect
Minimum viable	50–100	Noticeable improvement on narrow tasks; risk of overfitting.
Good	500–1,000	Reliable performance on your domain. Most projects target this.
Strong	1,000–10,000	High-quality, consistent behavior across edge cases.
Production-grade	10,000+	Near-expert level on complex, multi-faceted tasks.

Quality dominates quantity. 500 perfectly curated, diverse examples will outperform 50,000 noisy, repetitive ones every time. The single best place to spend an extra week is on data quality, not on hyperparameter sweeps.

The 2025–2026 cost landscape

Path	Approximate cost	What you need
OpenAI managed fine-tuning	$0.80–$25 per 1M training tokens	Just data + an API key.
Google Colab (free tier)	$0 (limited GPU time)	T4 GPU (~15 GB VRAM). Fine for QLoRA on 7B models.
RunPod / Lambda Labs / vast.ai	$0.30–$3.00 per GPU-hour	A100 / H100 rental. Pay-as-you-go.
Your own consumer GPU	$700–$2,000 upfront	RTX 3090 / 4090 with 24 GB VRAM.
Apple Silicon (MLX)	$0 if you already own the Mac	M2/M3/M4 Ultra with 64 GB+ unified memory.

Honest budget for a typical project

A first real fine-tuning project — Llama 3.1 8B with QLoRA on ~1,000 examples — costs roughly $0 (Colab free) to $10 (RunPod RTX 4090, ~3 hours). The expensive failure mode is not the GPU bill; it is the time you spend curating data, the iteration on hyperparameters, and the eval suite. Plan in days, not GPU-hours.

Code

Estimate OpenAI training cost from a JSONL file·python

import json

PRICE_PER_M_TOKENS = {
    "gpt-4.1-mini-2025-04-14": 0.80,
    "gpt-4o-mini-2024-07-18": 3.00,
    "gpt-4o-2024-08-06": 25.00,
}

def estimate_cost(jsonl_path: str, model: str, epochs: int = 3) -> float:
    chars = sum(len(line) for line in open(jsonl_path))
    tokens = chars // 4  # cheap approximation: 4 chars/token in English
    training_tokens = tokens * epochs
    cost = training_tokens / 1_000_000 * PRICE_PER_M_TOKENS[model]
    print(f"Lines: {sum(1 for _ in open(jsonl_path)):,}")
    print(f"Estimated tokens: ~{tokens:,}")
    print(f"Training tokens (× {epochs} epochs): ~{training_tokens:,}")
    print(f"Estimated cost on {model}: ~${cost:.2f}")
    return cost

estimate_cost("training_data.jsonl", "gpt-4.1-mini-2025-04-14", epochs=3)

Exercise

Pick a real or synthetic 100-example dataset and estimate the OpenAI fine-tuning cost on gpt-4.1-mini for 1, 3, and 5 epochs. Then estimate the same workload on RunPod (rent an A100 ~2 hours at $1.50/hr) for an open-source equivalent. Compare and write down which path you would actually pick and why.