CuneiScribe

Bridge a 4,000-year cultural gap — read and write in cuneiform.

CuneiScribe lets you interact with humanity's oldest writing system. Translate between English and Akkadian, convert to cuneiform Unicode signs, and render clay tablet images — with built-in confidence gating that tells you when results are unreliable.

$ cuneiscribe cuneiform "LUGAL dan-nu LUGAL KUR aš-šur"
𒈗 𒆗𒉡 𒈗 𒆳 𒀸𒋩

$ cuneiscribe classify "The computer sends an email"
Type:       modern
Confidence: 0.70
Warnings:
  - Contains 2 modern concept(s) with no direct Akkadian equivalent

All outputs are machine-generated approximations. Consult Assyriological expertise for research or public-facing use.

Showcase

Reading the past: Epic of Gilgamesh → English

Input (Akkadian):  ša naq-ba i-mu-ru i-šid ma-a-ti ša kul-la-ta i-du-u₂
Cuneiform:         𒊭 𒅘𒁀 𒄿𒈬𒊒 𒄿𒋃 𒈠𒀀𒋾 𒊭 𒆰𒆷𒋫 𒄿𒁺𒌑
English:           "What Naqba saw, the fortress of all the lands, which knows all of them."

Writing the past: Shakespeare → Cuneiform

Input (English):   "To be, or not to be, that is the question"
Akkadian:          šá-a-šú u la šá-a-šú šá-a-lu

Input (English):   "The king sent a letter to his brother"
Akkadian:          LUGAL a-na ŠEŠ-šu i-sap-ra
Cuneiform:         𒈗 𒀀𒈾 𒋀𒋗 𒄿𒉺𒅁𒊏

Confidence gating in action

$ cuneiscribe classify "The king rules the land"
Type:       short
Confidence: 0.85
Mode:       experience          ← Safe to render

$ cuneiscribe classify "Send me an email about the algorithm"
Type:       modern
Confidence: 0.70
Mode:       educational
Warnings:
  - Contains 2 modern concept(s) with no direct Akkadian equivalent
                                ← Warning: these concepts have no Akkadian equivalent

$ cuneiscribe classify "<script>alert(1)</script>"
Type:       anomalous
Confidence: 0.90
Warnings:
  - Input rejected              ← Blocked: anomalous input

Install

pip install cuneiscribe

Quick Start

CLI

# Convert transliteration to cuneiform
cuneiscribe cuneiform "LUGAL dan-nu"        # → 𒈗 𒆗𒉡

# Classify input before processing
cuneiscribe classify "The king rules"       # → short, 0.85, experience mode

# Render as clay tablet
cuneiscribe render "šar kiš-ša-ti" -o tablet.svg

# Full pipeline with confidence gating (requires model)
cuneiscribe craft "The king rules" --model models/byt5-base-akkadian --json

# Look up a sign
cuneiscribe info LUGAL                      # → 𒈗, U+12217

Python API

from cuneiscribe import CuneiScribe, classify

# Check input first
result = classify("I love pizza")
print(result.input_type)   # "short"
print(result.warnings)     # []

# Full pipeline with gating
tc = CuneiScribe(model_path="models/byt5-base-akkadian")
result = tc.craft("The mighty king")

print(result.akkadian)     # Transliteration
print(result.cuneiform)    # Unicode cuneiform
print(result.confidence)   # 0.0-1.0
print(result.suggestion)   # "render" / "render_with_caveat" / "fallback"
print(result.warnings)     # List of caveats

# No model needed for transliteration → cuneiform
result = tc.transliterate_and_render("LUGAL dan-nu", output_path="tablet.svg")

Web Demo

pip install cuneiscribe[serve]
python -m cuneiscribe.interfaces.demo --model models/byt5-base-akkadian --share

How It Works

User Input
    │
    ▼
Input Classifier ──→ anomalous? → REJECT
    │
    ▼
ByT5 Translation (En→Ak)
    │
    ▼
Output Validator ──→ unreliable? → FALLBACK (transliteration only)
    │
    ▼
Cuneiform Converter (14,240 mappings)
    │
    ▼
Tablet Renderer (SVG/PNG)
    │
    ▼
Result + Confidence + Warnings

The confidence gating pipeline ensures the system never confidently renders wrong cuneiform. When output quality is uncertain, it degrades gracefully to transliteration-only with a warning.

Features

Architecture

cuneiscribe/
├── pipeline/      ← Confidence gating (classifier + validator)
├── models/        ← ByT5 bidirectional translator
├── knowledge/     ← Sign tables + cuneiform converter
└── interfaces/    ← CLI, web demo, SVG renderer

Four decoupled layers. Swap the model, update sign tables, or add a dialect without breaking interfaces.

Limitations

• English→Akkadian produces approximate modern transliterations, not authentic ancient text
• Trained on Neo-Assyrian/Old Babylonian data; other Akkadian dialects (e.g., Old Assyrian commercial texts) may have substantially lower quality
• Even with 14,240 sign mappings and a 17K-lemma dictionary, domain-specific data matters more than model size
• Currently English-only

See ROADMAP.md for the engineering roadmap.

Citation

@inproceedings{cuneiscribe2026,
  title={CuneiScribe: Bridging a 4,000-Year Cultural Gap with Bidirectional Akkadian NMT and Cuneiform Rendering},
  author={Wang, Geoffrey},
  booktitle={Proceedings of the 4th Workshop on Cross-Cultural Considerations in NLP (C3NLP)},
  year={2026}
}

License

Apache 2.0