Cooklang Fine-Tuning Dataset

Training data for fine-tuning LLM models to convert recipes into Cooklang format.

The dataset is built from HelloFresh recipe pages scraped across multiple countries.

Structure

recipes/           Converted recipe pairs (.recipe → .cook) used as training data
inbox/             Raw scraped recipes not yet converted
labler/            Web UI for reviewing and correcting .recipe → .cook conversions
src/main.rs        HelloFresh scraper (Rust)
batch_convert.py   Batch conversion script using OpenAI Batch API

Data

recipes/ — Training Pairs

~17,700 recipe pairs across 11 countries (at, au, be, de, es, fr, gb, ie, it, nl, us). Each recipe has:

• .recipe — raw structured text scraped from HelloFresh (YAML frontmatter + ingredients + steps)
• .cook — the same recipe converted to Cooklang format

These pairs form the training data for fine-tuning.

inbox/ — Unconverted Recipes

~19,700 raw .recipe files from 4 additional countries (ca, ch, lu, nz) awaiting conversion.

Tools

Scraper (Rust)

Scrapes HelloFresh sitemaps and downloads recipes using cooklang-import.

cargo run -- --countries us,gb --output recipes
cargo run -- --list-countries

Requires cooklang-import binary (default path: ../cooklang-import/target/debug/cooklang-import).

Batch Converter (Python)

Converts .recipe files to .cook using the OpenAI Batch API with a fine-tuned model.

pip install -r requirements.txt

python batch_convert.py prepare   # Generate JSONL batch input
python batch_convert.py submit    # Upload and start batch job
python batch_convert.py collect   # Download results when complete

Requires OPENAI_API_KEY env var. Optionally set OPENAI_MODEL to override the default model.

Labler (Rust)

Web UI for reviewing and correcting .recipe → .cook conversions side-by-side. Three-panel layout with syntax-highlighted Cooklang editor, real-time parsing, and diff view.

cd labler
cargo run                      # All recipes (defaults to ../recipes)
cargo run -- ../recipes/us     # Only US recipes

See labler/README.md for details.

Validation Workflow

Multiple people can collaborate on reviewing recipes by splitting country folders between them.

Setup

Each reviewer picks a country folder to work on:

cd labler
cargo run -- ../recipes/us     # Reviewer A
cargo run -- ../recipes/gb     # Reviewer B

Review Process

1. Open a recipe and compare the original .recipe (left panel) with the Cooklang source (right panel)
2. Fix any conversion issues in the Cooklang editor — the middle panel shows a live preview
3. Use "Show Diff" to spot differences between the original and the rendered output
4. Once the recipe looks correct, add fine_tune_status: reviewed to the YAML frontmatter in the .cook file:

---
fine_tune_status: reviewed
---

Preheat oven to 425 degrees. Dice @potatoes{12%oz}...

5. Save (⌘S) and move to the next recipe (⌘])

Status Tracking

The fine_tune_status frontmatter field in .cook files tracks review progress:

Value	Meaning
(missing)	Not yet reviewed
reviewed	Human-verified and corrected

This lets the team see at a glance which recipes are done and which still need attention.

Recipe Format

Input (.recipe)

---
title: Recipe Name
description: ...
image: https://...
servings: 2
time required: 30m
nutrition:
  calories: 500 kcal
  ...
---

2 unit Onion
1 tablespoon Olive Oil
...

• Dice the onion.
• Heat olive oil in a pan.
...

Output (.cook)

Dice the @onion{2}. Heat @olive oil{1%tablespoon} in a #pan{}.

License

The recipe content is sourced from HelloFresh and remains their intellectual property. This dataset is provided for research and educational purposes only.