Toponym Context Classifier

DeBERTa-v3-large fine-tuned for discourse context classification of Ukrainian toponym mentions.

Part of #KyivNotKiev.

Performance

Model F1 Macro Accuracy
DeBERTa-v3-large (this) 0.857 +/- 0.013 0.901
XLM-RoBERTa-large 0.846 +/- 0.011 0.892
mDeBERTa-v3-base 0.807 +/- 0.007 0.864

11 context classes: war_conflict, academic_science, history, politics, sports, culture_arts, food_cuisine, travel_tourism, religion, business_economy, general_news.

Training

  • Corpus: 36,791 texts, 59 toponym pairs, 5 sources
  • Annotation: Claude Haiku 4.5 (validated: kappa=0.56-0.69 vs human consensus 86.2%)
  • Loss: class-weighted cross-entropy
  • Config: LR=1e-5, epochs=3, batch=16, fp16, seed=456
  • Benchmark: 21 runs (12 hyperparameter + 9 architecture comparison)

License

CC-BY-4.0

Downloads last month
13
Safetensors
Model size
0.4B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train KyivNotKiev/toponym-context-classifier