Fine-tune it — or just write a better prompt?
Scrub a shelf of cached, illustrative LoRA training curves — rank, learning rate, epochs — and let the break-even calculator name the exact request count where fine-tuning beats a better prompt.
Training loss (cached curve)
lossEval score over epochs (held-out)
evalWatch the split: training loss keeps falling while the held-out eval peaks and then decays past the overfit point — the checkpoint you ship is the peak, not the last epoch.
Tune vs prompt — the break-even
| One-time tuning — data prep $180 + 4.1 GPU-h × $2.40/h × rank factor 1.15 | $190 |
| Tuned request — 240 tokens on the compact base | $0.00 |
| Prompt request — 1800 few-shot tokens on a large routed model | $0.01 |
| At 100,000 requests: fine-tune $204 vs prompt $720 | tune wins here |
Cumulative cost vs volume (log scale, up to $36.0k)
prompt path fine-tune pathFine-tuning is normally a black-box decision made on vendor slides. Here you scrub a shelf of cached, clearly-illustrative training curves yourself — and the break-even calculator hands you a hard, named request count instead of a vibe.
What is LoRA fine-tuning?
LoRA (low-rank adaptation) trains small adapter matrices on top of a frozen base model instead of updating every weight. That makes tuning dramatically cheaper, and the adapters are tiny files you can swap per task — which is why it's the default way open-weight models get specialized.
When does fine-tuning beat writing a better prompt?
When volume amortizes the one-time training cost. A tuned model usually needs far fewer prompt tokens per request and can run on a smaller, cheaper base — so each request saves a little. Divide the all-in training cost by the per-request saving and you get the break-even request count where the tune starts winning.
Why does the eval score fall while the loss keeps dropping?
That's overfitting: past a certain epoch the model memorizes the training set, so training loss keeps improving while held-out eval quality peaks and then declines. It's why practitioners checkpoint every epoch and ship the peak — never automatically the last one.
This playable is part of the registry: its tool page
Next: will it fit on your GPU?
Can You Actually Run This? itemizes the VRAM bill — weights, KV cache at your context length, and overhead — against real GPU memory sizes.