Tune or Prompt?

In progress

The live, in-browser demo for Tune or Prompt? is being built. The full technical teardown is on the right →

A study with a conclusion: exactly when fine-tuning a small model earns its keep — and when it doesn’t.

Impact

4-way

comparison

~N req

break-even

cost crossover

Messy product input → clean ad {hook, headline, body, cta}. For quality alone a prompted frontier model already wins, so the honest reason to fine-tune is cost at scale. This proves the crossover with numbers, not vibes.

The problem

Everyone reaches for fine-tuning to improve quality. But a prompted frontier model already generates great structured ads. The real question isn’t “can I fine-tune,” it’s “when does fine-tuning a small model become the right call?”

The approach

Establish both baselines (prompted-frontier ceiling, prompted-small floor), LoRA-tune a small model, and measure JSON validity, field accuracy, and refusal-correctness — then a four-way comparison on quality AND cost, with a clear crossover.

Architecture

Task dataset

2k–10k clean pairs

PROCESS

Baselines

prompted frontier + small

MODEL

LoRA / QLoRA SFT

Qwen3-8B · 1 GPU

MODEL

DPO preference tune

optional depth

OUT

Cost-crossover decision

a number, not a vibe

·Establish the quality ceiling (frontier) and the flaky floor (prompted small) first.
·A tuned small model matches frontier quality — the win is cost per call at scale.
·The deliverable is the request volume where fine-tuning beats prompting.

How it was built

Baselines

Make the gap real

✓Prompted frontier — the ceiling
✓Prompted small — the flaky floor

Phase 1

SFT

✓2k–10k clean examples
✓LoRA / QLoRA on a small base
✓Eval JSON validity, field accuracy

Phase 2

DPO

✓Preference pairs, train with DPO
✓Show the gain over SFT

Phase 3

Decision write-up

✓Four-way quality + cost comparison
✓The crossover: pays off past ~N requests

Stack

HuggingFace TRLLoRA / QLoRADPOAxolotlQwen3-8B

Source

GitHub ↗Decision write-up

Want this on your team?

hello@miske.dev ← back to work