Sample bundle: ML experiment trail (suggested types: Experiment, Model, Submission, Lesson)

[topprismdata]

Summary

The three published sample bundles (ga4, stackoverflow, crypto_bitcoin) are all data catalog style — they describe external resources (BigQuery tables, Stack Overflow tags, Bitcoin data). I built a fourth style: an ML experiment trail (see bundle at /tmp/s6e2-okf-bundle/).

Suggested new type values for ML domain

The SPEC (§4.1) explicitly says types are free, but it would help consumers (visualize, future readers) if a few ML-canonical types were documented:

• Experiment — a single training run / iteration (R0, v15f, etc.)
• Model — a trained model (e.g., CatBoost BAG L1, OOF 0.9555)
• Submission — a file submitted to a Kaggle LB
• Lesson — a discovered or validated skill/principle
• Competition — a Kaggle competition as a whole (timeline container)
• Bundle — entry point for the whole bundle (top-level index.md)

Why a sample bundle helps

The current GA4 bundle shows OKF for static metadata. An experiment-trail bundle would demonstrate:

1. Cross-link graph shape — 11 concepts / 23 edges in my mini-bundle, dense and navigable.
2. Type-driven color coding in visualize — already works; would just need ML types documented.
3. Reference to non-file resources — resource: https://github.com/.../pull/5 (PR/issue links) and resource: file://... patterns.
4. Time-evolving knowledge — timestamp field shines here: each run is dated, the dashboard.md can be regenerated from latest timestamp.

My S6E2 mini-bundle (12 concepts, 6 types)

The mini-bundle I built covers an AutoGluon rerun of Kaggle Playground Series S6E2 (Heart Disease). It demonstrates:

• 1 Bundle (root)
• 3 Index (subdirectory hubs)
• 3 Experiment (R0 baseline, R0a bug, R0b fix)
• 3 Model (CatBoost BAG L1, LightGBM BAG L1, Weighted Ensemble L3)
• 2 Submission (0/1 vs probability)
• 4 Lesson (4 validated/refined skills)

I'd be happy to donate this as a sample bundle under okf/samples/ if the maintainers want.

Related

• Issue Producer: convert any OpenAPI 3.x spec into an OKF v0.1 bundle #56 (OpenAPI→OKF) is in the same family — adding more producer-types.
• Issue Spec: add recommended provenance fields — author, publisher, published, license #52 (provenance fields) is adjacent — my Submission concepts use resource: file:// and could benefit from published_at once that lands.

[topprismdata]

Status update (2026-06-21)

I shipped two related PRs to topprismdata/cultivating-ml-agent since opening this issue:

• PR Update README.md #11: S6E2 LR stacker no-gain lesson — lesson concept in OKF format
• PR Add EntryLinks support #12: S6E6 CV-LB gap lesson (stacker OOF overfitting) — new decision rule (gap > 0.01 = stop stacker)

Both follow the suggested ML-canonical type values from this issue (Experiment, Model, Submission, Lesson).

Question for maintainer

1. Is there interest in merging any of these into the official okf/bundles/ samples?
2. The three issues Sample bundle: ML experiment trail (suggested types: Experiment, Model, Submission, Lesson) #60, Sample bundle: ML experiment trail (suggested types: Experiment, Model, Submission, Lesson) #71, Sample bundle: ML experiment trail (suggested types: Experiment, Model, Submission, Lesson) #72 all propose "Sample bundle: ML experiment trail" — should they be consolidated, with Sample bundle: ML experiment trail (suggested types: Experiment, Model, Submission, Lesson) #71 (in cultivating-ml-agent) as the canonical one?

Happy to provide a PR against the OKF repo if helpful.

Filed by automated follow-up from cultivating-ml-agent pipeline.