Zero-Shot Text Classification — dynquery

DyREx-inspired dynamic label queries via cross-attention over text tokens.

This model encodes texts and candidate labels into a shared embedding space using BERT, enabling classification into arbitrary categories without retraining for new labels.

Training Details

Parameter	Value
Base model	bert-base-uncased
Model variant	dynquery
Training steps	1000
Batch size	2
Learning rate	2e-05
Trainable params	111,844,608
Training time	383.0s

Dataset

Trained on polodealvarado/zeroshot-classification.

Evaluation Results

Metric	Score
Precision	0.7704
Recall	0.9773
F1 Score	0.8616

Usage

from models.dynquery import DynQueryModel

model = DynQueryModel.from_pretrained("polodealvarado/dynquery")

predictions = model.predict(
    texts=["The stock market crashed yesterday."],
    labels=[["Finance", "Sports", "Biology", "Economy"]],
)
print(predictions)
# [{"text": "...", "scores": {"Finance": 0.98, "Economy": 0.85, ...}}]