← Back

📑 Contents

Enveda CASMI 2026Current stateWorking rulesResource constraintIntended first milestone

Enveda CASMI 2026


Working repository for the Kaggle competition [Enveda CASMI 2026 - Molecule ID From Mass Spectra](https://www.kaggle.com/competitions/enveda-CASMI26-molecule-id-mass-spectra).


Current state


  • Competition review and execution plan: [EXECUTION-PLAN.md](EXECUTION-PLAN.md)
  • Live task tracker: [TODO.md](TODO.md)
  • Official Kaggle page snapshots and public baseline research: [research/](research/)
  • Visible schema/test fixtures: [data/](data/)
  • Implementation scaffold: src/, tests/, configs/, manifests/, experiments/, notebooks/, reports/
  • No model training or competition submission has been run yet.

  • Working rules


    1. The scorer is molecule-level MRR@25, using tautomer-canonicalized RDKit InChIKey14 identity.

    2. Group all validation splits by molecule identity. Do not leak spectra or derived features across splits.

    3. Treat the downloaded test.parquet as a schema/plumbing fixture only; it is replaced by hidden data at scoring time.

    4. Keep raw competition data and external databases out of public git and do not redistribute them.

    5. Every experiment records its configuration, data/checkpoint hashes, split manifest, metrics and runtime.


    Resource constraint


    The development machine has approximately 11 GiB RAM. Do not load the full training Parquet into pandas or build unbounded dense indexes locally. Use PyArrow row-group streaming, selected columns, compact metadata tables, memory-mapped/bounded indexes and sampled validation. Large training and final inference workloads belong in Kaggle notebooks or another approved compute environment.


    Intended first milestone


    Build a reproducible scorer-compatible validation harness and a molecule-level spectral retrieval baseline before adding databases or generative models.