msc_thesis_nuextract3_stage2_best_all

This is Stage-2 Best All POIs (Food-Constrained) fine-tuned NuExtract3 Model.

Training Parameters: Stage-1:

• Dataset Size: 6092 weakly labelled data
• 5 Epochs (1902 Steps)
• Effective Batch Size: 16 (8 gradient accumulation steps * 2 batch size)
• Learning Rate: 2e-4
• Linear Warmup across 15 steps
• Weight Decay: 0.01
• Evaluate every 150 steps, early stopping if all validation scores do not increase over the best score over 2 evaluations

Stage-2:

• Dataset Size: 500 manully labelled data
• 10 Epochs (2500 Steps)
• Effective Batch Size: 2 (2 gradient accumulation steps * 1 batch size)
• Learning Rate: 3e-5
• Evaluate every 125 steps, early stopping if all validation scores do not increase over the best score over 2 evaluations

Best Stage-1 Occured at step 300, and best All POIs (Food-Constrained) Score occured at step 375

<b>Evaluation Scores on Validation and Test Dataset:</b><br> Both Validation and Test are 200-sized manually labelled dataset