Early-stopping state
Validate mode, patience and score inputs. Keep best epoch separate from stop epoch and make a stopped run stay stopped.
← Selected projects
LEARNING / REPRODUCIBILITY TOOLS
KAPT Reproducibility Lab
Turn a local model study into small, inspectable tools: validation selects the checkpoint; aligned predictions define the ensemble; evidence defines the claim.
- Organized
- 2026.10.14
- Upstream
- KANO · MIT ↗
- Public scope
- Learning tools & synthetic fixtures
01 / ENGINEERING STUDY
Make the evaluation contract explicit
Small components with clear inputs, state and failure conditions.
The local project combines the third-party KANO baseline with KAPT study code. Its evaluation helpers needed clearer boundaries: a model checkpoint is not a prediction file, averaged metrics are not an ensemble metric, and the epoch with the best validation score is not necessarily the epoch when training stops.
Prediction-level ensembles
Align rows by sample ID, check labels and prediction ranges, then evaluate the averaged predictions. Keep threshold choice on validation data.
Prompt tensor contracts
Task prompts, node-feature pooling and residual injection with explicit shapes. Pooling uses node features and graph IDs; it does not consume graph edges.
- 01
Validation scores
Mode · patience · min delta
- 02
Selected checkpoint
Best epoch · explicit state
- 03
Aligned predictions
Sample ID · consistent labels
- 04
Held-out evaluation
Fixed rule · stated limits
An original reading diagram for the public tools, not a claim that the full KANO/KAPT training pipeline has been reproduced.
02 / VALIDATION REPLAY
Best epoch ≠ stopped epoch
Synthetic validation curve · No model training or benchmark result.
- Current epoch
- 8
- Best epoch
- 5
- Without improvement
- 3 / 3
- Stop epoch
- 8
Best scoreStopping point
Epoch data & decisions
| Epoch | Score | Best epoch | Decision |
|---|
A score must improve by more than min delta to reset patience. Training stops after patience consecutive non-improving epochs. Restore the selected checkpoint before held-out evaluation; a later test score must not change the selection.
03 / SOURCE & LIMITATIONS
Read the implementation and its limits
01 / prompts.pyShapes, pooling & residual injection ↗02 / early_stopping.pySelection state & strict improvement ↗03 / ensemble.pySample IDs & prediction metrics ↗
KANO is credited to its original authors. The upstream MIT notice is preserved where applicable. The replay and reading diagram on this page are newly prepared synthetic learning material.