T5-Feature-Based-NQ: Feature-Based Transfer Learning on Natural Questions ShortQA

Model Overview

• Model Architecture: T5-base
• Training Type: Feature-based transfer learning (frozen encoder, trainable decoder)
• Task: Question Answering (Short-form answers)
• Dataset: lucadiliello/naturalquestionsshortqa (80/20 train/test split)
• Evaluation Metrics: Exact Match (EM), F1-score
• Pretrained Checkpoint: mrm8488/t5-base-finetuned-squadv2
• Framework: Hugging Face Transformers + PyTorch

Evaluation Results

Results are after 15 epochs of feature-based fine-tuning on GPU with frozen encoder.