Open AI Datasets & Corpora
This page provides technical documentation and dataset specifications for computer vision engineers, computational linguists, and machine learning researchers accessing our Hugging Face repositories.
Raw PDF Images and Multimodal Dataset Utility
Although our archives consist primarily of raw PDF page images, high-resolution document scans, and uncurated audio, they are fully usable as primary datasets. Modern multimodal vision-language models, document parsing algorithms, and OCR computer vision encoders rely directly on high-resolution page images. Raw PDF scans preserve 100% of authentic orthography, historical fonts, and page structures without OCR transcription error.
Simple Structure and No Formal Dataset Cleaning Plans
We intentionally keep the archive structure simple so our small team can focus on collection and public domain scanning. The Foundation has no plans to structure, annotate, or clean these archives into polished benchmark datasets. Any data cleaning, tokenization, filtering, or instruction formatting must be performed by the developers and researchers who download them.
Official Dataset Repositories & Hub Handles
You can inspect or download raw text corpora directly via Python using the Hugging Face datasets library via load_dataset("hmar-heritage") or by accessing our official repositories below:
| Repository / Hub | Platform | Direct Link |
|---|---|---|
| hmar heritage datasets hub | Hugging Face | huggingface.co/hmar-heritage |
| hmar heritage github organization | GitHub Org | github.com/hmar-heritage |
Read more about this archival initiative under our hmar corpus archival project brief.
Open Licensing and Commercial Derivative Works
Using our raw archives to train language models, computer vision algorithms, or software applications is 100% free. Reselling or paywalling our raw uncleaned archives as standalone products is prohibited. However, if a developer or company performs significant engineering work—such as cleaning, restructuring, annotating, alignment, or instruction-tuning the raw data into a refined dataset product—they are fully permitted to license or sell their restructured derivative work.