anaskhan1/urdu-factcheck-xlmr-adapted-100-per-class

UrduFactCheck Adapted XLM-R

This model starts from the counterfactual MT Urdu-FEVER XLM-R checkpoint and is further adapted on a small balanced UrduFactCheck split with 100 examples per class.

Labels:

• 0: SUPPORTS
• 1: REFUTES
• 2: NOT_ENOUGH_INFO

Main reported held-out UrduFactCheck result:

• Accuracy: 77.78%
• Macro F1: 0.7729
• SUPPORTS F1: 0.7101
• REFUTES F1: 0.6154
• NOT_ENOUGH_INFO F1: 0.9934

This is the final adapted model from the research project. For code, diagnostics, and project documentation, see:

https://github.com/Anas-Khaan/CrossLingual-Hallucination-NLI