# Dataset sources for common/data_prep.py. # Copy this file to configs/sources.yaml (git-ignored is recommended) and edit the paths. # # Each source is a set of Parquet files with one image column and one text column, as published # by most Hugging Face OCR datasets (download them with `huggingface-cli download --repo-type dataset`). # # files glob of Parquet files; {split} is replaced by each key in `splits`. # Relative paths are resolved from the repository root. # splits map of file split name -> split name written to prepared/canonical/.jsonl. # Map a source to `eval` to keep it for evaluation only. # image_col column holding image bytes (or {"bytes": ...}) default: image # text_col column holding the target text / Markdown default: text # level page (full page -> Markdown) or line (one text line) default: page # license recorded on every record so licences travel with the data # language optional ISO code, written to olmOCR front matter sources: pages: files: data/raw/my-page-dataset/data/{split}-*.parquet splits: {train: train, test: test} text_col: markdown level: page license: CC-BY-4.0 lines: files: data/raw/my-line-dataset/data/{split}-*.parquet splits: {train: train, validation: validation, test: test} text_col: text level: line license: CC-BY-4.0 benchmark: files: data/raw/my-benchmark/data/train-*.parquet splits: {train: eval} # evaluation only: never trained on text_col: markdown level: page license: Apache-2.0