oracc-parser 0.1.2

Download and parse ORACC cuneiform text projects into ML-ready formats

oracc-parser

A Python tool to download and parse ORACC cuneiform text projects into machine-learning-ready formats (JSONL, CSV, pandas DataFrames).

Features

• Download — Fetch project ZIPs directly from ORACC or Zenodo
• Parse — Convert raw ORACC JSON into structured data
• Export — Save datasets as JSONL, CSV, or pandas DataFrames
• Configure — Control handling of broken signs and POS masking using RunConfig

Installation

git clone https://github.com/shaharspencer/oracc-parser.git
cd oracc-parser
pip install -e ".[dev]"

Getting Started — Notebooks

The easiest way to explore oracc-parser is through the interactive notebooks. Start with notebook 01 — it downloads all the data you need from Zenodo automatically.

Notebook	What you'll learn
01_quickstart.ipynb	Download the dataset → parse a project from pre-processed CSVs → explore transliterations, translations, and metadata → export
02_reference_data.ipynb	Browse all projects in the dataset, query catalogues, explore bundled reference data (provenance, periods, sign list, POS tags)
03_configure_and_export.ipynb	All RunConfig options — word-level and sign-level break filtering, POS masking — combining multiple projects and exporting datasets
04_oracc_json_processing.ipynb	Advanced: understand the raw ORACC JSON structure, the JSON → TabletRecord → CSV pipeline, and how to download and parse projects not in the dataset

pip install oracc-parser[notebooks]
jupyter notebook notebooks/

Quick Example

from oracc_parser import parse_project, RunConfig, get_full_flat_table

# Parse 5 tablets from SAA 01 (Neo-Assyrian royal letters)
records = parse_project("saao/saa01", config=RunConfig(limit=5))

# Get a flat DataFrame — no nesting, ready for analysis
df = get_full_flat_table(records)
df.to_json("dataset.jsonl", orient="records", lines=True)

Configuration

You can customize the parsing process using RunConfig:

from oracc_parser import parse_project, RunConfig

records = parse_project("saao/saa01", config=RunConfig(
    limit=10,
    max_break_fraction=0.5,   # word-level: drop words that are >50% broken
    drop_missing=True,        # sign-level: drop [x] signs from Unicode output
    drop_damaged=False,       # sign-level: keep ⸢x⸣ signs in Unicode output
    mask_pos=["PN", "DN"],    # replace personal/divine names with tag
))

Two independent levels of break filtering

RunConfig provides two distinct ways to handle damaged or missing text, operating at different granularities and affecting different outputs:

Parameter	Level	Affects	How it works
max_break_fraction	Word	Transliteration, normalization, lemmatization	Each word has a break_perc (fraction of its signs that are broken). Words exceeding this threshold are replaced with X. Default 1.0 keeps all words.
drop_missing	Sign	Unicode cuneiform only	Drops individual signs marked [x] (completely lost).
drop_damaged	Sign	Unicode cuneiform only	Drops individual signs marked ⸢x⸣ (partially legible).

Note: Because word-level and sign-level filtering use different thresholds and different granularities, the text outputs and the Unicode cuneiform output are not necessarily aligned. A word kept in the transliteration (because its average damage is below max_break_fraction) may still have individual signs dropped from the Unicode output if drop_missing / drop_damaged are enabled.

Other options

Parameter	Default	Description
limit	None	Only parse the first N texts (useful for testing)
keep_word_segmentation	True	Preserve word boundaries in Unicode cuneiform output
mask_pos	[]	Replace words of certain POS tags with the tag name
languages	["Akkadian"]	Which languages to include when downloading projects
use_cache	True	Use cached results if available

All reference data is bundled with the package, so you don't need to configure external paths unless you are customizing oracc_parser.settings.

CLI

oracc-parser download --project saao/saa01
oracc-parser parse --project saao/saa01 --limit 5 --format jsonl --output saa01.jsonl

Heavy Data (Zenodo)

Large data files (ORACC ZIPs, cached translations, Pleiades data) are on Zenodo:

python scripts/download_zenodo_data.py

Running Tests

pytest tests/ -v     # 98 tests

Known Limitations

• Chronology: Period-to-year normalization is optimized for the 1st Millennium BCE.
• Language: Parsing is primarily validated on Akkadian projects.

License

MIT — see LICENSE.

Credits

Based on code by Niek Veldhuis (Compass) and adapted for the BEn Project.