This model is a fine-tuned version of microsoft/deberta-v3-base for binary disinformation detection. It classifies news articles as either credible (0) or disinformation (1).

Model Details

Base model: microsoft/deberta-v3-base
Task: Binary text classification (credible vs. disinformation)
Language: English
Training framework: PyTorch + Transformers

Training Hyperparameters

Parameter	Value
Learning rate	1e-05
Batch size (train)	16
Batch size (eval)	16
Epochs	5
Weight decay	0.01
Warmup ratio	0.06
FP16	True
Max sequence length	512
Seed	42
Eval steps	100
Best model selection	binary_f1_pos

Evaluation Results

Overall (Test Set)

Metric	Value
Binary F1 (positive)	0.9041
Macro F1	0.9342
Accuracy	0.948
AUC-ROC	0.9864
Precision	0.9223
Recall	0.9485

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("pjait/deberta-v3-base-disinfo-task1-binary")
model = AutoModelForSequenceClassification.from_pretrained("pjait/deberta-v3-base-disinfo-task1-binary")

text = "Your article text here..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    logits = model(**inputs).logits
    probability = torch.sigmoid(logits).item()
    prediction = "disinformation" if probability >= 0.5 else "credible"

print(f"Prediction: {prediction} (probability: {probability:.4f})")

Training Data

The model was trained on the about 5k articles dataset for Task 1 (binary classification), which contains news articles annotated by multiple annotators for credibility assessment.

Limitations

The model is trained on English-language articles only.
Performance may vary on domains or topics not represented in the training data.
The model should be used as a tool to assist human judgment, not as a sole decision-maker.

Citation

If you use this model, please cite the paper (to do, currently paper under review).

pjait/deberta-disinfo-detection

DeBERTa-v3-base for Disinformation Detection (Binary Classification)