MD Mizanur Rahman

Work · Agricultural vision

Coconut leaf benchmark and audit trail

  • Demo ready
  • Synthetic demonstration
  • Companion to coauthored research

A benchmark scaffold where no number can exist without a saved predictions file behind it, built as a companion to the coauthored coconut-leaf article.

Repository on GitHub

Problem

The coauthored article reports its accuracy in two slightly different ways. A companion benchmark should make that impossible: every metric regenerates from saved predictions and the denominators come from the manifest.

Contribution

Relationship to research: Companion to coauthored research. New companion benchmark for the coauthored coconut-leaf article. The article's dataset, split records and permissions have not been obtained; its accuracy figures are not restated or used as a target.

Account owner’s role: Commissioned, reviewed and published the implementation. The code was produced with AI-assisted generation on commission and reviewed before release; it is not the original code of any cited publication.

What was built

  • Manifest-driven dataset loader
  • Small CNN and transfer model
  • report command that refuses without predictions or on hash mismatch

Method

Metrics are recomputed only from predictions.csv with its manifest hash verified.

Data and access

No dataset bundled; a checklist of what to request from the original authors is documented.

Evaluation protocol

Accuracy, macro F1, per-class metrics, confusion matrix, calibration, split counts from the manifest.

Results

No measured result is published for this project. Demonstration outputs from the repository’s own synthetic fixtures are labelled as such in the repository and are not presented here as results.

Data and experiment lineage

Run 20260913T014010+0000-demo-small_cnn, mode demo, synthetic fixture. Chain verification: every hash recomputed and matched (export sha256 1f4d40f367e2)

  1. dataset

    source
    synthetic-fixture-v1 v1
    files
    80
    file-list hash
    3e34a9e101a0…

  2. split

    artefact
    split_manifest.json
    sha256
    c7bfa4f8ca2d…
    split counts
    test 16 · train 48 · val 16
    depends on
    files_hash 3e34a9e101a0…

  3. config

    artefact
    config.resolved.yaml
    sha256
    604330beef7d…

  4. run

    artefact
    manifest.yaml
    sha256
    c9ab87e4dd22…
    seed
    42
    status
    completed
    checkpoint
    not committed (hash recorded)
    depends on
    split_manifest_hash c7bfa4f8ca2d…; configuration_hash 604330beef7d…

  5. predictions

    artefact
    predictions.csv
    sha256
    911ee6710237…
    prediction rows
    16 (test)
    depends on
    split_manifest_hash c7bfa4f8ca2d…; configuration_hash 604330beef7d…

Every sha256 was recomputed from the file at export time. Fixture runs are not benchmark results. Placeholder class names are used until the dataset owners supply the real definitions; the repository’s report command refuses to compute any metric unless the predictions file hashes back to this chain.

Limitations

  • Original dataset and split records not obtained; class definitions are placeholders until supplied.

Reproducibility

Audit chain: dataset, split, config, run, predictions, each hashed.

Repository metadata: Python, MIT, last push 2026-09-13, 0 stars.

Source publication and links