m-E5 WeightedSum + MeanPooling Polarization Classifier (Pro)
This model uses multilingual-e5-large with WeightedSum + MeanPooling architecture for polarization classification.
Architecture
- Base Model: intfloat/multilingual-e5-large
- Layer Pooling: Learnable weighted sum of last 6 layers
- Sequence Pooling: Mean pooling over tokens
- Head: Dropout(0.1) → Linear(1024, 2)
- Output: 2-way logits (polarization classification)
Training
- Loss: CrossEntropyLoss
- Sampling: Random (default)
- Frozen Layers: Bottom 12 layers
- Languages: 22 languages
- Max Length: 256
Prompting Format
Uses m-E5 instruction format:
Instruct: Classify the polarization of this {Language} text.
Query: {text}
Languages
amh, arb, ben, deu, eng, fas, hau, hin, ita, khm, mya, nep, ori, pan, pol, rus, spa, swa, tel, tur, urd, zho
