Minimal Reproducibility Package

Minimal public repository to rerun the core IMERG correctability benchmark on a compact smoke dataset.

This package is intentionally narrow:

• one temporal scale: 3 h
• one compact derived dataset: 345,742 rows
• three representative IMERG cells, one per province: Alicante, Castellon, Valencia
• core model ladder only: M0, M1, M2, M3, M3b
• evaluation and event diagnostics included
• deep learning (M4) and full-domain 1 h / 6 h experiments excluded

The goal is not to redistribute the full study inputs. The goal is to provide a GitHub-friendly repository that reruns the main experimental logic end to end with small, versionable files.

Repository Layout

paper_8_imerg_correctability_minimal/
  data/
    manifests/
      extreme_thresholds_by_scale.csv
      split_assignments_3h.parquet
    processed/
      model_dataset_extremes_3h.parquet
  scripts/
    06_baseline_climatology.py
    07_quantile_mapping.py
    08_ml_tabular.py
    08_evaluate.py
    09_ml_two_stage.py
    11_build_diagnostics.py
    12_build_smoke_figure.py
    verify_outputs.py
    utils/
  results/
    figures/
  run_pipeline.py
  requirements.txt

What This Reproduces

The pipeline runs:

1. M1 climatological correction
2. M2 quantile mapping
3. M3 LightGBM tabular correction
4. M3b two-stage LightGBM correction
5. out-of-sample benchmark evaluation
6. event-based timing and tolerance diagnostics
7. one compact summary figure for the smoke benchmark

Main outputs are written under results/:

• results/metrics/benchmark_metrics_all_models.csv
• results/metrics/correctability_metrics_vs_m0.csv
• results/diagnostics/event_peak_summary.csv
• results/diagnostics/tolerance_sensitivity.csv
• results/figures/figure_smoke_correctability.png

Install

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Tested with Python 3.13.2.

Run

python run_pipeline.py

If you want to skip the event diagnostics or the summary figure:

python run_pipeline.py --skip-diagnostics
python run_pipeline.py --skip-figure

Data Notes

• The repository includes a small derived benchmark subset only.
• Raw AVAMET data and raw IMERG HDF5 files are not redistributed here.
• The included parquet files are sufficient to rerun the public smoke benchmark without any external downloads.

Scope Notes

• This is a reproducibility package for the experimental scaffold, not the full publication archive.
• The numerical outputs from this smoke subset should not be read as the headline paper results.
• The full study used a larger domain, additional scales, more diagnostics, and supplementary experiments not included here.

License

MIT. See LICENSE.