deberta-v3-base-figurative

DeBERTa-v3-base fine-tuned on VUA20 + MAGPIE + FLUTE for 4-class figurative language detection (literal / idiom / metaphor / simile).

Training

Fine-tuned on a combination of:

• VUA20 — VU Amsterdam Metaphor Corpus (metaphor)
• MAGPIE — idiom dataset
• FLUTE — figurative language understanding

Labels: literal (0), idiom (1), metaphor (2), simile (3) Epochs: up to 20 (early stopping, patience=2) | Hardware: 1× NVIDIA A100 40 GB

Intended use

Frozen English teacher in the CLKD pipeline for Plains Cree figurative language detection. The teacher produces soft-label distributions on English translations; these are distilled into a student encoder operating on Cree text.

Results

Evaluated on a held-out split of VUA20 + MAGPIE + FLUTE.

Metric	Value
Macro F1	0.8699
Literal F1	0.8397
Idiom F1	0.9729
Metaphor F1	0.7623
Simile F1	0.9048

Citation

If you use this model, please cite the associated thesis/paper (TBD).

Data

Training data includes:

• Bloomfield (1934) Plains Cree Texts (scraped and sentence-aligned)
• EdTeKLA Indigenous Languages Corpora