DistilBERT IMDb Sentiment Classifier

Model Description

• Base model: distilbert-base-uncased-finetuned-sst-2-english
• Task: Sentiment Analysis (binary classification: positive / negative)
• Dataset: IMDb (50,000 labeled movie reviews)
• Max input length: 256 tokens
• Labels:
  • 0: Negative
  • 1: Positive

Training Details

• Dataset split: IMDb train (25k) / test (25k)
• Optimizer: AdamW (lr=5e-5)
• Batch size: 64 (gradient accumulation steps = 2)
• Epochs: 15
• Weight decay: 0.01
• Warmup ratio: 0.1
• Evaluation: Performed at the end of each epoch
• Hardware: GPU with optional FP16 training
• Other settings: gradient accumulation, fused AdamW optimizer, multiple dataloader workers

Results

• Loss: 0.4525
• Accuracy: 0.91396
• Precision: 0.9028
• Recall: 0.9278
• F1 Score: 0.9151