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

Downloads last month
10
Safetensors
Model size
0.6B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support