convaiinnovations/medgemma-ecg-training-metrics

MedGemma ECG Training Metrics

Training metrics and visualization plots from fine-tuning MedGemma-4B on ECG datasets.

Overview

This repository contains training metrics, loss curves, and performance visualizations from fine-tuning Google's MedGemma-4B-it model on ECG interpretation tasks using the PTB-XL subset of the ECGInstruct dataset.

Training Details

Model: google/medgemma-4b-it (fine-tuned with LoRA)
Dataset: PULSE-ECG/ECGInstruct (PTB-XL subset)
Infrastructure: AIRAWAT (C-DAC) - 8x NVIDIA A100 40GB GPUs
Training Duration: ~16.5 hours (2 epochs)

Final Metrics

Metric	Value
Token Accuracy	89.62%
Training Loss	0.99
Entropy	0.985
Total Tokens	103,301,284

Visualization Plots

Training Dashboard

Combined view of all training metrics including loss, accuracy, learning rate, gradient norm, and entropy.

Individual Plots

Plot	Description
	Training loss over epochs with smoothed trend line
	Token accuracy (train & eval) over time
	Cosine learning rate schedule
	Gradient norm stability
	Model entropy (confidence) over training

Training Configuration

# LoRA Settings
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05

# Training Hyperparameters
epochs: 2
learning_rate: 2e-4
batch_size: 192 (effective)
optimizer: AdamW (fused)
lr_scheduler: cosine
precision: bfloat16
gradient_checkpointing: true

Key Observations

1. Loss Convergence: Training loss decreased smoothly from ~10 to ~1.6
2. Accuracy Improvement: Token accuracy improved from 50% (random) to 89.6%
3. Stable Training: Gradient norms remained stable (0.7-0.9)
4. Entropy Reduction: Model became more confident over training

Related Resources

• Fine-tuned Model: convaiinnovations/medgemma-4b-ptbxl
• Base Model: google/medgemma-4b-it
• Dataset: PULSE-ECG/ECGInstruct

Acknowledgments

• Infrastructure: AIRAWAT AI Innovation Challenge by C-DAC
• Base Model: Google MedGemma team
• Dataset: PULSE-ECG team

License

Apache 2.0