kaggle-submission-flow

name	kaggle-submission-flow
description	This skill should be used when the user asks to "submit to kaggle", "kaggle submit", "push submission", "submit experiment", "サブミット", "sub して", "提出して", "kaggle に提出", "check submission results", "結果確認", "スコア確認", "public score", "LBスコア記録", "submission failed", "サブミット失敗", or wants to run the Kaggle submission pipeline (deps push, model upload, kernel push, result check, troubleshooting).

Kaggle Submission Flow

Guide the full Kaggle submission pipeline: metadata setup, dependency push, model upload, kernel push, result verification, and score recording.

Pipeline Overview

Step	Task Command	Description
1. Metadata setup	task setup-kaggle-metadata EXP=expXXX	Generate kernel-metadata.json and code.ipynb
2. Full submission	task submit-kaggle EXP=expXXX	Push deps + upload model + push kernel
3. Result check	uv run kaggle competitions submissions -c $COMPETITION_NAME	Check public score via API
4. Record results	Update README.md front matter	Write public_lb score

Submission Scenarios

Scenario A: Full Pipeline (First Submission or Model Updated)

Run the full pipeline when the model artifacts have changed:

task submit-kaggle EXP=expXXX

This executes three steps sequentially:

1. push-kaggle-deps — Push dependency wheels to Kaggle (skipped if PUSH_DEPS=false)
2. push-kaggle-models — Upload model artifacts (can take several minutes for large models)
3. sleep 60 + push-kaggle-sub — Wait for model processing, then push submission kernel

Execute task submit-kaggle with the Bash tool's run_in_background: true parameter, since model upload takes several minutes and blocking the conversation prevents other work.

Scenario B: Model Already Uploaded (Re-submit Only)

When the model is already uploaded and only the kernel code or metadata changed:

cd models/<exp>/submission && uv run kaggle k push

This only pushes the submission kernel and completes in seconds.

Scenario C: Dependencies Updated

When deps/requirements.txt has changed, ensure deps are pushed first:

task submit-kaggle EXP=expXXX PUSH_DEPS=true

PUSH_DEPS=true is the default. Set PUSH_DEPS=false to skip deps push when dependencies haven't changed.

Scenario D: Switch Experiment on Same Kernel

All experiments share the same submission kernel name (auto-shortened from competition name). To submit a different experiment, regenerate metadata and push:

task setup-kaggle-metadata EXP=expXXX
task submit-kaggle EXP=expXXX PUSH_DEPS=false

Metadata Setup

Metadata generation is typically handled during experiment creation (task new-exp), but can be regenerated:

task setup-kaggle-metadata EXP=expXXX

This runs four commands:

1. deps-metadata — Generate deps/kernel-metadata.json
2. deps-code — Generate deps/code.ipynb
3. submission-code — Generate models/<exp>/submission/code.ipynb
4. submission-metadata — Generate models/<exp>/submission/kernel-metadata.json

Key metadata details:

• Kernel name: Shared across experiments, auto-shortened from competition name to fit Kaggle's 50-char title limit
• GPU: Enabled by default (enable_gpu: true)
• Model sources: Points to {username}/{comp}-models/other/{exp}/1
• Deps kernel: Bundled as a kernel source for offline pip install

Checking Submission Results

API Check

uv run kaggle competitions submissions -c $COMPETITION_NAME

Output columns: fileName, date, status, publicScore, privateScore.

Wait for status to become SubmissionStatus.COMPLETE before reading scores. While the kernel is still running, status shows SubmissionStatus.PENDING or similar.

Kernel Execution Log

If a submission fails, check the kernel output on Kaggle. Common failures:

• ModuleNotFoundError — missing package in deps
• Model file not found — model not uploaded or wrong experiment name
• GPU OOM — model too large for allocated GPU

Recording Results

After confirming the public score, update the backlog task:

backlog task edit TASK-N --append-notes "Public LB: XX.XX"

Troubleshooting

Issue	Cause	Fix
400 Bad Request on kernel push	Title exceeds 50 chars	Run task setup-kaggle-metadata (auto-shortens)
ModuleNotFoundError	Missing package in deps	Add to deps/requirements.txt, push deps
Model file not found	Model not uploaded or wrong exp	Run task push-kaggle-models EXP=expXXX
GPU OOM	Model too large for GPU	Adjust enable_gpu in metadata
Kernel timeout	Inference too slow	Optimize batch size or model