StyleECU-es
StyleECU-es is a style embedding model for Spanish, obtained by fine-tuning mStyleDistance on SynthSTEL-ES, a purpose-built Spanish contrastive dataset of 51,400 triplets covering 71 stylistic dimensions.
Model Description
StyleECU-es specializes the mStyleDistance embedding space toward stylistic phenomena most relevant to Spanish, including dialectal variation (voseo/tuteo), expressive morphology, syntactic complexity, and digital style.
Training
- Base model:
StyleDistance/mstyledistance - Training objective: TripletLoss (contrastive learning)
- Dataset: style-anon/SynthSTEL-ES
- Training size: 51,400 triplets
- Epochs: 2
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("style-anon/StyleECU-es")
embeddings = model.encode(["Your text here"])
Evaluation
Evaluated on PAN author profiling tasks (Spanish):
| Task | Base (mStyleDistance) | StyleECU-es | Δ |
|---|---|---|---|
| PAN 2018 – Gender prediction | baseline | +3 pp | +3 pp |
| PAN 2021 – Hate speech spreaders | 0.70 | 0.81 | +11 pp |
Authors
Citation
If you use this model, please cite:
Paper under review. Citation will be updated upon publication.