Store the current run, then change values and store another run. The comparison table stays in this browser until cleared.
| Run | Model | Dataset | Best val acc. | Lowest val loss | Duration | Learning rate | Parameters |
|---|---|---|---|---|---|---|---|
| No stored runs. | |||||||
Why Keep a Model Training Log?
A model training log creates a reliable history of every experiment. Machine learning work often involves many small configuration changes. Without structured records, it becomes difficult to explain why one run performed better. A useful log connects model architecture, dataset versions, preprocessing choices, hyperparameters, hardware, timing, and evaluation results. This makes experiments easier to reproduce and review.
Track More Than Accuracy
Training accuracy alone rarely describes model quality. Validation loss, precision, recall, F1 score, ROC-AUC, regression errors, or task-specific metrics may reveal different weaknesses. Epoch-level tracking also shows whether training remains stable. A widening difference between training and validation performance can indicate overfitting. Sudden loss spikes may expose learning-rate problems, corrupted batches, unstable gradients, or data issues.
Improve Reproducibility
Reproducibility depends on more than a random seed. Record framework versions, dependency versions, dataset releases, Git commits, environment details, and configuration snapshots. Hardware can also influence training speed and numerical behavior. Checkpoint information helps teams identify exactly which saved model produced an evaluation result. Good records reduce duplicated work when experiments are repeated weeks later.
Compare Experiments Efficiently
Experiment comparison should focus on both quality and cost. A slightly better score may not justify several times more compute, memory, or training time. Use this generator to compare model variants, optimizers, learning rates, datasets, and resource requirements. Exported CSV data supports deeper analysis, while PDF reports are convenient for reviews and documentation. Consistent logs turn isolated training runs into useful evidence for future model decisions.
Turn Logs Into Better Decisions
Training logs also improve communication across engineering and research teams. Reviewers can see what changed, what remained constant, and which result justified a new model version. Recording failures is equally valuable because failed experiments can prevent repeated mistakes. Add concise conclusions after every completed run. Note promising follow-up ideas and unresolved questions. Over time, these records create a practical knowledge base for selecting architectures, tuning parameters, estimating compute needs, and planning future experiments.
Common Training Metrics Reference
| Metric | Typical use | Interpretation |
|---|---|---|
| Loss | Most training tasks | Lower values usually indicate better objective optimization. |
| Accuracy | Balanced classification | Shows the fraction of predictions classified correctly. |
| Precision | False-positive-sensitive tasks | Measures correctness among predicted positive samples. |
| Recall | False-negative-sensitive tasks | Measures how many actual positives were detected. |
| F1 score | Imbalanced classification | Balances precision and recall with one score. |
| RMSE | Regression | Measures prediction error in the target's units. |
| IoU / Dice | Segmentation | Measures overlap between predicted and reference regions. |
Frequently Asked Questions
What should be recorded for each epoch?
Record training loss, validation loss, important metrics, learning rate, duration, and notable events.
Why does validation loss matter?
Validation loss estimates generalization and can reveal overfitting before training accuracy shows a problem.
How do I identify the best checkpoint?
Choose a selection metric that matches the task, then save the checkpoint with the strongest validation result.
What is a generalization gap?
It is the difference between training and validation performance. A large gap can indicate overfitting.
Why record the Git commit?
The commit identifies the exact code version used during training, improving reproducibility and debugging.
Can this tool handle regression or segmentation?
Yes. Use the general evaluation fields and custom metric column for task-specific measurements.
Are training records uploaded anywhere?
This page works locally in the browser. Export actions create files directly from entered information.