boeliebamba/isc-helwigii-polyphony-v1

I.S.C. Helwigii Polyphony Disambiguation Model v1.0

Fine-tuned bert-base-multilingual-cased for cuneiform sign reading disambiguation (polyphony resolution).

Model Description

This model predicts the correct reading of a polyphonous cuneiform sign given its textual context. It was trained on 37,139 tablets from the CuneiML v1.2 dataset containing 2.8M signs, covering 19,993 mythological texts with 31,287 motif instances across 10 myth categories and 16 motif types.

Intended Use

• Primary: Disambiguate cuneiform sign readings in transliterations
• Secondary: Assist in automated transliteration and translation pipelines
• Research: Digital Assyriology, computational cuneiform studies

Training Data

• Dataset: CuneiML v1.2 (37,139 tablets, 2.8M signs)
• Splits: 500K train / 50K validation / 500K test examples
• Labels: 57,752 unique sign readings (Sumerian + Akkadian)
• Context window: ±3 signs (7 total tokens including target)
• Max sequence length: 128 tokens

Model Architecture

• Base model: bert-base-multilingual-cased (179M parameters)
• Task: Sequence classification
• Fine-tuning: Full model fine-tuning (3 epochs)
• Batch size: 32
• Learning rate: 2e-5
• Weight decay: 0.01
• Warmup: 10% of training steps

Performance

Metric	Value
Validation Accuracy	{accuracy}
Validation F1 (weighted)	{f1}
Validation Loss	{loss}

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("boeliebamba/isc-helwigii-polyphony-v1")
model = AutoModelForSequenceClassification.from_pretrained("boeliebamba/isc-helwigii-polyphony-v1")

# Example: disambiguate sign 𒀭 (AN) in context
context = "d en-lil2 [SEP] AN"
inputs = tokenizer(context, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    logits = model(**inputs).logits
predicted_id = logits.argmax(-1).item()

Citation

@misc{helwigii-polyphony-v1,
  author = {Helwig, M.},
  title = {I.S.C. Helwigii Polyphony Disambiguation Model v1.0},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/boeliebamba/isc-helwigii-polyphony-v1}
}

License

MIT License — see LICENSE for details.