[finetune.lab]./start

v0 · private_beta

┌──────────────────────────────────────────────────────────────────┐
│  $ fine-tune --idea="your one-liner" --model=auto --dataset=auto │
│  > booting research agents ...                          [ READY ]│
└──────────────────────────────────────────────────────────────────┘

Fine-tune.a.model
for.your.idea_

Describe what you want. Get the right open model, the right dataset, and a ready-to-run training script — in five minutes.

./start_project ./see_examples

hugging_faceunslothmlxlora · qlorashadow_eval

product

Three things most fine-tune projects skip — and fail on.

01[feature]

Dialog-driven SPEC

A senior ML engineer asks the right questions. You answer. In five minutes you walk out with a clean SPEC — task, data shape, success metric, constraints.

02[feature]

Auto research

We scan Hugging Face for the best open-source model and the right dataset for your problem. No vibe-based picks.

03[feature]

Ready-to-run script

Unsloth for GPU, MLX for Apple Silicon. Plus a shadow-eval harness so you know your fine-tune actually beats the baseline before you ship.

examples

Real fine-tunes from a one-line idea.

Product recommender

Fine-tune a small Gemma to pick the right ceramic adhesive from a catalog of 791 SKUs.

./p/kerakoll-selector-v2

Crypto signal classifier

Classify bullish / bearish signals from crypto-Twitter. 56k samples, Gemma4-E4B.

./p/cryptogemma

UI element detector

Detect clickable UI elements from screenshots. ScreenSpot-v2 benchmark, 2-stage LoRA.

./p/clickgemma-v2

method

Seven phases. Hard gates. Auto-drop on failure.

No vanity metrics. No vibe-checks. Every phase produces an artifact — or kills the project.

[0]

Framing

5-7 questions → SPEC.md

[1]

Baseline

closed-model baseline to beat

[2]

Dataset

open dataset + validation split

[3]

Sanity

small LoRA, must beat baseline

[4]

Full train

full fine-tune on sanity-approved config

[5]

Shadow eval

blind eval on held-out prod samples

[6]

Ship or drop

hard auto-drop on no baseline win

From idea to fine-tuned model
Without the 3-day research detour.

./start_project ./how_it_works