Free playground · no account

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.

Illustrative simulation — authored dataset, no live training runs
LoRA rank
Learning rate

Training loss (cached curve)

loss
3.0 loss0epoch 0epoch 3

Eval score over epochs (held-out)

eval
best 87 @ ep 310030epoch 0epoch 3

Watch 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 $720tune wins here

Cumulative cost vs volume (log scale, up to $36.0k)

prompt path fine-tune path
break-even ~26,900$36.0k$01,0005M requests
Fine-tuning starts winning at ~26,900 requestsIllustrative numbers — the crossover logic is the takeaway, not the digits.
The Demystify reveal

Fine-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.