joelbarmettler/md-reheader-dataset

md-reheader Dataset

Training data for md-reheader, a model that restores heading hierarchy in markdown documents.

Dataset Description

Each example is a ChatML conversation with three messages:

1. System: Instructions for heading level prediction
2. User: A markdown document with all headings flattened to # (level 1) and body text stripped to first 128 + last 128 tokens per section
3. Assistant: The correct headings with proper # prefixes (e.g., ## Methods, ### Data)

Example

{
  "messages": [
    {"role": "system", "content": "You are a markdown document structure expert..."},
    {"role": "user", "content": "# Introduction\n\nFirst 128 tokens...\n\n# Background\n\n..."},
    {"role": "assistant", "content": "# Introduction\n## Background"}
  ],
  "metadata": {
    "source": "goodwiki",
    "token_count": 1632,
    "heading_count": 9,
    "max_depth": 3
  }
}

Dataset Structure

Split	Examples	Notes
train	196,949	Includes 2-8x oversampling of deep-structure documents
validation	7,321
test	7,321

Metadata fields

• source: github_code or goodwiki
• token_count: Token count of the stripped document (Qwen3 tokenizer)
• heading_count: Number of headings in the document
• max_depth: Maximum heading depth (1-6)
• min_depth: Minimum heading depth
• max_level_gap: Largest gap between adjacent heading levels

Sources

codeparrot/github-code (~105k documents)

Markdown files from GitHub repositories, loaded from the codeparrot/github-code dataset. Filtered by .md/.markdown file extension. Bucket-targeted download ensures representation across document lengths (< 4k, 4k-8k, 8k-16k, 16k-32k characters).

Split by repository name to prevent data leakage.

euirim/goodwiki (~45k documents)

High-quality Wikipedia articles from euirim/goodwiki. Article titles prepended as # {title} since Wikipedia's title serves as the H1 heading.

Split by article title to prevent data leakage.

Processing Pipeline

1. Download raw markdown documents from HuggingFace
2. Filter by character length (300-150k), heading count (3-120), distinct levels (≥2), token count (100-31k)
3. Split by source key (repo name / article title) into 90% train / 5% val / 5% test
4. Oversample deep-structure documents in the training split (depth 4: 2x, depth 5: 4x, depth 6: 8x)
5. Strip body text to first 128 + last 128 tokens per section
6. Flatten all headings to # (level 1)
7. Truncate documents exceeding 7,500 tokens
8. Format as ChatML conversations

Heading Level Distribution (training set)

Level	Count	Percentage
H1	211,673	8.5%
H2	911,437	36.8%
H3	922,695	37.3%
H4	335,052	13.5%
H5	78,761	3.2%
H6	16,345	0.7%

License

This dataset is released under CC-BY-SA 4.0 because it contains derivative content from Wikipedia.

• Wikipedia content: CC-BY-SA 4.0 by Wikimedia contributors
• GitHub code files: various permissive licenses (MIT, Apache 2.0, BSD, etc.) via codeparrot/github-code
• GoodWiki dataset: CC-BY-SA 4.0 by Euirim Choi

Author

Built by Joel Barmettler.

Citation

@software{barmettler2026mdreheader,
  author = {Barmettler, Joel},
  title = {md-reheader: Restoring Heading Hierarchy in Markdown Documents},
  year = {2026},
  url = {https://github.com/joelbarmettlerUZH/md-reheader}
}