ciscoriordan/dilemma-model

Dilemma Model Weights

Character-level transformer for Greek lemmatization, used as the neural fallback in the Dilemma lemmatizer.

Model Details

• Architecture: Encoder-decoder transformer (character-level)
• Parameters: 4.2M
• d_model: 256
• Attention heads: 4
• Layers: 3 (encoder and decoder)
• Feed-forward dim: 512
• Vocabulary: 381 characters (Greek polytonic + special tokens)
• Training data: 3.4M form-lemma pairs from Wiktionary inflection tables
• Multi-task heads: POS tagging (10 tags), nominal inflection (45 labels), verbal inflection (69 labels)

Files

File	Size	Description
model.pt	~16 MB	PyTorch checkpoint (weights + config)
encoder.onnx	~7 MB	ONNX encoder for lightweight inference
decoder_step.onnx	~10 MB	ONNX decoder for lightweight inference
vocab.json	~9 KB	Character vocabulary (char2id / id2char mappings)

Usage

This model is used automatically by the Dilemma library as a fallback for forms not found in the 12.3M-entry lookup table or resolved by rule-based morphological analysis. Only about 5% of Greek words reach the transformer.

pip install dilemma

from dilemma import Dilemma
d = Dilemma()

# The transformer is invoked automatically when needed
d.lemmatize("ἐποιήσαντο", lang="grc")  # -> ποιέω

ONNX vs PyTorch

For inference, ONNX Runtime (50 MB install) and PyTorch (2 GB install) produce identical results. The ONNX files are provided for environments where a lighter dependency is preferred. PyTorch is only needed for training.

Training

Trained from scratch in minutes on a single GPU using the train.py script in the Dilemma repository:

python train.py
python export_onnx.py  # Export to ONNX format

License

MIT