CarbonResearchLabs/fullerene-data

Fullerene training data: manifest and profile

The training data manifest of Fullerene, the model family of Carbon-OCR, an OCR engine in C++ and CUDA by Carbon Research. It is the data rec-latin and det of fullerene-3M train on: the Latin set, 2.77 M parameters in its four models. The models document explains every Fullerene model, layer by layer.

This repository holds no third-party data. It records where every file of the training set comes from, its size and SHA-256, the checksum its source publishes, and its license, and it measures what is inside each source. The data stays on its original hosts, under its own licenses; the scripts here fetch it and check it.

The training set, measured 2026-10-27

• Sources: 21 of 21
• Size: 142.6 GB fetched, against 156.9 GB planned
• Files: 14,786
  • 14,772 match the checksum or size their source publishes
  • 14 are re-encoded slices of larger files
  • 0 failed a check

The full report, including what the data holds and how it compares with what other OCR engines trained on, is reports/phase1/REPORT.md.

The benchmarks

Evaluation sets only, never trained or tuned on: 41 of the 51 benchmark sources beyond the training set's own evaluation sets (page parsing, mixed-image OCR, forms and receipts, tables, math, and per-language sets), 3.2 GB. Each benchmark's manifest records the exact commit it was fetched at (revision), since a benchmark's scores compare only at the same revision. The training set's manifests predate that field and are checked against each file's checksum.

Files

path	what
manifests/phase1/<source>.json	the training set, per source: URL, license, allowed use, and every file with its size, SHA-256 and upstream checksum
manifests/bench/<source>.json	the same for each benchmark, with the commit it was fetched at
reports/phase1/REPORT.md, summary.json	verification, contents, comparison with reference engines, gaps
reports/phase1/profile/<source>.json	per source: counts, character sets, scripts, line geometry, layout labels, PDF text layers, font coverage
scripts/	sources.py (the sources, slices and sets), fetch.py (download and verify: --set phase1, the training set, or --set bench), profile.py, report.py

Fetch the same set again

huggingface-cli download CarbonResearchLabs/fullerene-data --repo-type dataset --local-dir fullerene-data
python -m venv .venv && .venv/bin/pip install huggingface_hub pyarrow h5py pillow pypdfium2 fonttools brotli
.venv/bin/python fullerene-data/scripts/fetch.py --root carbon-data               # the training set
.venv/bin/python fullerene-data/scripts/fetch.py --set bench --root carbon-data   # the benchmarks

fetch.py downloads each source from its own host, checks every file against the checksum the source publishes, and writes its manifests beside it (scripts/manifests/), to compare with the ones here; a rerun resumes.

License

• This manifest and report: CC BY 4.0.
• Each data source: its own license, recorded per source in its manifest (license, license_url, use).
  • use T: may train weights that are shipped.
  • use E: evaluation only.
  • use check: the license is to be confirmed before training.
• Fullerene's weights: the Carbon-OCR OpenRAIL-AMS license (LICENSE-WEIGHTS.md), CarbonResearchLabs/fullerene-3M. Carbon-OCR's code: Apache 2.0.