Fixing T5 Fine-Tuning Bugs for Medical Diagnosis

This notebook documents the debugging process of fine-tuning T5-small for medical diagnosis tasks. What started as a simple text-to-text classification project turned into a deep dive into tokenizer handling, generation bugs, and training instabilities. Disclaimer: this is for educational purposes only and not for actual medical diagnosis.

Problem

After training, the model only output "True" or repeated inputs. Despite a great loss curve, the outputs were useless.

Method

I approached this by:

• Isolating tokenizer and data collator issues
• Testing task prefix impact (critical for T5!)
• Comparing fresh vs continued training
• Tuning generation parameters (e.g. beam search, repetition penalty)

Results

Final loss: 0.009; accuracy on our small, hand-crafted test set was 100%, likely due to the limited dataset and clear-cut labels (e.g., Pneumonia, Myocardial infarction). This should not be interpreted as clinical performance.