← Back
📝 Notes / kaggle-kernel / haideptry-v32-replica / README

📑 Contents

Haideptry V32 faithful Kaggle replica (visible-test bank included)Included behaviorProvenanceSafety and interpretationRun

Haideptry V32 faithful Kaggle replica (visible-test bank included)


This is the private baseline recreation of the public notebook haideptry/0-350-casmi-2026-v32-ensemble.


Included behavior


This replica restores the retrieved notebook's full inference logic and all eight upstream attached datasets, including haideptry/casmi26-v33-champion-artifacts and its test_verified_isomers_with_ms2.parquet artifact. That bank has 15,147 rows covering 386 molecule IDs; the original run log says its positive-fitness subset loaded for 354 visible IDs. The engine joins the bank by molecule_id, interleaves bank structures into later candidate ranks, and uses generated isomers when the bank is small. This is query-specific visible-test behavior and is included here specifically for a faithful reproduction, not endorsed as transferable inference.


Provenance


  • Upstream notebook source preserved at upstream/0-350-casmi-2026-v32-ensemble.ipynb.
  • Upstream metadata preserved at upstream/kernel-metadata.json.
  • Pulled source notebook SHA-256: edea80e9091ebfe2f4d2690e33c73c4cd6e19952cb5d340c1826e69efff90580.
  • The four executable modules are extracted verbatim from source notebook %%writefile cells; hashes are in replica-manifest.json.
  • Upstream slug/metadata title calls this V32; notebook's original header calls the method V33/0.365+ target. No saved cell outputs or execution counts are present, and the score shown in discussion (0.337) is not independently evidenced by this notebook/log.

    • Safety and interpretation


  • The visible-test-ID bank contains candidates scored using experimental test spectra. Its provenance/scoring calculation and competition-rule eligibility have not been independently audited. Treat the run as a query-specific/transductive reproduction; do not infer hidden performance or label-leakage facts from it.
  • The bank is attached intentionally for source fidelity; the user's selected baseline is this bank-included replica.
  • A completed notebook run and valid CSV structure are not a leaderboard score. Outputs must be independently checked against runtime IDs; no competition submission is made by this package.
  • Sources include a mixture of licenses. In particular COCONUT metadata declares CC BY 4.0; ranker/model datasets declare CC0 in Kaggle metadata; ChEBI/LIPID MAPS declares CC BY-NC-SA 4.0; offline RDKit lists license other; check source terms for intended use.

    • Run


    Kaggle kernel slug is in kernel-metadata.json and is private. Attachments mirror the currently retrievable upstream metadata. Push/run only this private kernel for reproduction; do not submit the CSV to the competition unless separately decided.