LightOnOCR-mix-0126

LightOnOCR-mix-0126 is a large-scale OCR training dataset built via distillation: a strong vision–language model is prompted to produce naturally ordered full-page transcriptions (Markdown with LaTeX math spans and HTML tables) from rendered document pages. The dataset is designed as supervision for end-to-end OCR / document-understanding models that aim to output clean, human-readable text in a consistent format.

This repository releases the PDFA-derived annotated subset used in the LightOnOCR-2-1B-0126 training mixture. We do not distribute the source PDFs; only the text targets (and associated metadata) are provided. The source documents can be recovered from pixparse/pdfa-eng-wds. Quick links:

• 📃 Paper: arxiv
• 🤗 Model: lightonai/LightOnOCR-2-1B
• 🛠️ Code: (add link)
• 📄 Source pdfs: pixparse/pdfa-eng-wds

What’s in the dataset

Each row corresponds to a single page with:

• a text transcription in natural reading order
• markup for structure (headers, lists, tables) and math (LaTeX inside math spans)
• lightweight metadata emitted by our normalization / validation pipeline

Teacher model and supervision format

Annotations were generated using a state-of-the-art VLM teacher (see the LightOnOCR-2 paper for details). Compared to earlier generations, the newer teacher yields cleaner mathematical notation and fewer formatting artifacts, which improves downstream OCR model stability.

The target format is Markdown with:

• LaTeX math restricted to math spans
• standardized image placeholders when relevant without bounding box information (see lightonai/LightOnOCR-bbox-mix-0126 if required)
• tables represented in minimal html formatting without styling

Normalization and cleanup

Raw VLM transcriptions often contain superficial inconsistencies (e.g., stray Markdown fences, variable placeholders, templated “empty page” messages, occasional format drift). While these rarely affect human readability, they increase target entropy and hurt deduplication and learning stability. We therefore apply a unified normalization pipeline before release, including:

• Text sanitation: remove spurious Markdown code fences / ticks; harmonize whitespace.
• Deduplication / filtering: compute hashes on normalized text and filter common failure modes such as loop-like repetition.
• LaTeX validation: enforce formatting invariants (LaTeX confined to math spans) and optionally check KaTeX compatibility.
  The conversion step emits structured metadata that enables reproducible filtering rules (e.g., exclude unresolved references or KaTeX-incompatible samples).

Bounding boxes (separate dataset)

During distillation, the teacher occasionally emitted figure bounding box coordinates even when not explicitly prompted. Manual inspection of a sampled subset indicated these boxes are highly accurate. To keep the base OCR objective focused on transcription, these coordinate traces are removed from the main OCR targets, but retained as a separate supervision signal for image localization.

This extracted and reformatted subset is released as:

• lightonai/LightOnOCR-bbox-mix-0126

Data format

• key: unique identifier for the source pdf
• page_idx: source page of the pdf document used for transcription
• content: normalized transcription target (Markdown with LaTeX math spans and HTML tables)
• metadata:
  • element_counts:
    • formulas: number of latex-delimited formulas in the transcription
    • images: number of image placeholders in the transcription
    • tables: number of HTML tables in the transcription
  • token_length: token size of the transcription (using LightOnOCR-2-1B-0126 tokenizer model)

Note: This dataset release does not include the source PDF files or rendered page images.

Intended use

• Training / fine-tuning end-to-end OCR VLMs that output natural reading order text
• Studying OCR robustness to scientific markup (math, references, structured text)
• Benchmarking formatting stability / normalization techniques

Not intended for:

• reconstructing or redistributing the original PDFs
• high-stakes applications without further validation

Limitations

• Targets are model-generated and may contain occasional hallucinations or formatting errors, especially on extremely complex layouts.
• Language coverage is strongest for European-language content; performance may vary on scripts not well represented in the underlying sources.

Licensing / Terms of Use

## PDFA-derived portion (from PDFA / SafeDocs / CC-MAIN-2021-31-PDF-UNTRUNCATED)

This dataset includes annotations derived from PDF documents originating from the PDFA / SafeDocs corpus (CC-MAIN-2021-31-PDF-UNTRUNCATED), in the same spirit as the pixparse/pdfa-eng-wds release. The PDFA-derived portion of this dataset is provided subject to the upstream conditions: users must comply with Common Crawl’s license and terms of use and the Digital Corpora project’s Terms of Use.

Citation

@misc{lightonocr2_2026,
  title        = {LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR},
  author       = {Said Taghadouini and Adrien Cavaill\`{e}s and Baptiste Aubertin},
  year         = {2026},
  howpublished = {\url{https://arxiv.org/pdf/2601.14251}}
}