Debk/Ben_NER_xlm-roberta-large_mod_v6

Bengali NER Model

Model Description

This model is fine-tuned XLM-RoBERTa for Bengali Named Entity Recognition with proper tokenization handling.

Architecture Details

• Base model: xlm-roberta-base
• Task: Named Entity Recognition
• Language: Bengali
• Tokenization: XLM-RoBERTa tokenizer with proper Bengali word handling
• Architecture:
  • BiLSTM layers: 2
  • Hidden size: 256
  • CRF decoder
• Training Date: 2025-01-13

Tokenization

• Uses XLM-RoBERTa tokenizer
• Handles Bengali word splitting
• Maintains proper label alignment for split words
• Uses word_ids for token-to-word mapping

Performance Metrics

• final_train_loss: 1.2787
• final_val_loss: 3.5568
• best_val_loss: 3.0780
• epochs_trained: 5
• training_args: {'epochs': 5, 'batch_size': 16, 'learning_rate': 2e-05, 'warmup_ratio': 0.1, 'gradient_accumulation_steps': 2, 'max_length': 91, 'lstm_hidden_size': 256, 'lstm_layers': 2, 'lstm_dropout': 0.2}