Home/Docs/Kaggle kernels
Active Kaggle scope

The current kernel index.

DueCare's active Kaggle submission centers on three script kernels. Historical appendix kernels are archived and should not be treated as the current reviewer path. Source on GitHub; canonical index at kaggle/_INDEX.md.

Active reviewer kernels

These are the maintained paths for the current submission: exploration, focused demo, and quantitative proof. They share the same Gemma runtime and harness primitives instead of duplicating model-loading or harness logic.

01 · app DueCare App Interactive workbench for chat, harness comparison, bulk review, extraction, search safety, anonymization, grading, and layer introspection.
Kaggle → GitHub
02 · live demo DueCare Live Demo Focused, presentation-friendly path for showing the shared Gemma runtime, layered harness behavior, and reviewer-facing demo flow. Hosts the recording-grade pitch deck at /start and /slides.
Kaggle → GitHub
03 · quantitative proof DueCare Fine-tuning and Evaluation Benchmark, synthetic-data, optional LoRA training, checkpoint/resume, combined judging, and report/export artifacts.
Kaggle → GitHub

Shared contracts

The kernels are intentionally narrow wrappers around shared runtime and harness components. Exact catalog sizes are runtime metadata; static prose uses stable capability language to avoid drift.

What should stay true

  • 01, 02, and A-00 use the shared Gemma 4 runtime for inference model loads.
  • The default offline proof path uses the same chat harness primitives as the exploration workbench comparison page.
  • A-00 saves benchmark outputs, training artifacts, checkpoints, activity logs, and final reports under Kaggle working directories.
  • Archived appendix kernels should not be revived for the current submission without an explicit new decision.