Training Reports

Automatically generated reports for classifier fine-tuning runs. Each report compiles metadata, hyperparameters, model metrics, system resource usage, and the per-epoch training log from data recorded by Weights & Biases.

Available reports

Run ID	Date	Best Val Loss	Best Epoch	GPU	Duration
hummingbird_classifier_20260430_193523	2026-04-30	0.0804	7	NVIDIA L40S	7m 45s

Generate a new report

After a training run completes (and the wandb data has been saved alongside the model checkpoint), generate a report from the project root:

python docs/scripts/generate_training_report.py hummingbird_classifier_<run_id>

This produces:

• docs/training-reports/<run_id>.md — the report page
• docs/assets/images/training_reports/ — figures (loss, accuracy, F1, GPU/CPU metrics)

The report is linked here automatically — add a row to the table above to include it in the navigation.

Report contents

Each report includes:

Section	Content
Header	Run ID, W&B link, git commit, SLURM job ID
Compute Resources	Host, GPU model/memory/cores, CPU, RAM, partition
Hyperparameters	Requested vs. completed epochs, batch size, LR, seed
Best Result	Metrics at the epoch with lowest validation loss
Final Epoch Metrics	All metrics from the last completed epoch
Model Metrics	Loss/accuracy curves, precision/recall/F1, learning rate
System Metrics	GPU utilization/memory/temp, CPU, RAM, disk usage
Per-Epoch Log	Raw training log lines for each epoch