Subodh_MFND_mdeberta_v3

This model is a LoRA fine-tuned version of microsoft/mdeberta-v3-base for multilingual fake news detection (Bangla, English, Hindi, Spanish).
Final evaluation set results:

• Accuracy: 95.41%
• F1: 0.95
• (Precision/Recall can be filled in if you have them.)

Model description

• Privacy-preserved, multi-lingual fake news detection.
• Fine-tuned with LoRA adapters (r=8, α=16, dropout=0.1).
• Batch size: 8, Epochs: 3, Learning rate: 2e-4.

Intended uses & limitations

• Intended for research and production on multilingual fake news detection tasks.
• Works on Bangla, English, Hindi, and Spanish news content.
• Not intended for languages outside the fine-tuning set.

Training and evaluation data

• Dataset: Custom multilingual fake news corpus (Bangla, English, Hindi, Spanish)
• Supervised classification (fake/real)

Training procedure

Training hyperparameters

• learning_rate: 0.0002
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: AdamW
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
0.4942	1.0	9375	0.4617	0.7785	0.7776
0.4948	2.0	18750	0.4684	0.7591	0.7424
0.4892	3.0	28125	0.4376	0.7702	0.7569
Final Test	-	-	-	0.9541	0.95

Framework versions

• PEFT 0.17.1
• Transformers 4.56.1
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.22.0