DeFactoX

From Fragments to Facts: Curriculum-Driven DPO for Hindi News Veracity Explanations

Pulkit Bansal · Raghvendra Kumar · Shakti Singh · Adam Jatowt · Sriparna Saha

Abstract

We introduce DeFactoX, a framework for Hindi news veracity prediction and explanation generation. The approach combines Curriculum Learning with Direct Preference Optimization (DPO), enhanced through Hin-DPO, which incorporates two key signals: Actuality (factual correctness) and Finesse (hallucination stability). This leads to more reliable, coherent, and human-aligned explanations for misinformation detection in low-resource languages.

Method Overview

DeFactoX pipeline: dataset construction → curriculum learning → Hin-DPO training.

Key Contributions

• First framework for Hindi veracity explanation generation
• Novel Hin-DPO loss with Actuality + Finesse
• Curriculum learning for progressive alignment
• Large-scale synthetic Hindi preference dataset

Results

Our method consistently improves explanation quality across multiple models (mT5, LLaMA, Mistral), achieving higher ROUGE, METEOR, and BERTScore compared to standard DPO and SFT baselines.

Example Outputs

Below are examples comparing human explanations with model-generated outputs.

Citation

@article{defactox2026,
  title={From Fragments to Facts: A Curriculum-Driven DPO Approach for Hindi News Veracity Explanations},
  author={Bansal, Pulkit and Kumar, Raghvendra and Singh, Shakti and Jatowt, Adam and Saha, Sriparna},
  year={2026}
}