Track Kaggle Experiments with W&B — Even Without Internet Access

The Problem

Kaggle Notebooks disable internet access for competition submissions.

This means you can't push metrics to Weights & Biases in real time — wandb.log() calls silently fail, and your experiment tracking is gone.

I built a CLI tool to fix this.

kaggle-wandb-sync

pip install kaggle-wandb-sync

PyPI: https://pypi.org/project/kaggle-wandb-sync/
GitHub: https://github.com/yasumorishima/kaggle-wandb-sync

The idea: use WANDB_MODE=offline to log everything locally inside Kaggle, then download and sync after the run finishes.

Notebook runs offline → download output → wandb sync → W&B cloud

Notebook Setup

Add two lines before importing wandb:

import os
os.environ['WANDB_MODE'] = 'offline'   # must be before import wandb
os.environ['WANDB_PROJECT'] = 'my-project'

import wandb
wandb.init(name="my-run")
# ... your training code ...
wandb.log({"loss": 0.1, "accuracy": 0.95})
wandb.finish()

The offline run is saved to Kaggle's output directory automatically.

Sync with One Command

After the notebook finishes running:

export WANDB_API_KEY=your_api_key

# Push notebook, wait for completion, download output, sync to W&B
kaggle-wandb-sync run my-competition/

Or step by step:

kaggle-wandb-sync push   my-competition/           # push notebook
kaggle-wandb-sync poll   username/my-competition   # wait for COMPLETE
kaggle-wandb-sync output username/my-competition   # download output
kaggle-wandb-sync sync   ./kaggle_output           # sync to W&B

Log Submission Scores

After submitting to the Kaggle leaderboard, record the score directly to your W&B run:

kaggle-wandb-sync score https://wandb.ai/me/my-project/runs/abc123 \
  --tm-score 0.17 \
  --rank 779

This updates the run summary with tm_score, kaggle_rank, and automatically sets submitted: True (as of v0.1.4).

For other metrics, use --metric:

kaggle-wandb-sync score https://wandb.ai/.../runs/abc123 \
  --metric auc=0.95 \
  --metric logloss=0.32 \
  --rank 42

GitHub Actions Integration

For a fully automated pipeline, add this workflow:

name: Kaggle W&B Sync

on:
  workflow_dispatch:
    inputs:
      notebook_dir:
        description: "Notebook directory"
        required: true

jobs:
  sync:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install
        run: pip install kaggle-wandb-sync

      - name: Set up Kaggle credentials
        run: |
          mkdir -p ~/.kaggle
          echo '${{ secrets.KAGGLE_API_TOKEN }}' > ~/.kaggle/kaggle.json
          chmod 600 ~/.kaggle/kaggle.json

      - name: Run pipeline
        env:
          WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
        run: kaggle-wandb-sync run ${{ inputs.notebook_dir }}

Trigger it from the GitHub Actions UI after your notebook finishes.

Real-world Example: Stanford RNA 3D Folding 2

I used this to track experiments in Stanford RNA 3D Folding 2.

The notebook uses template matching — finding similar RNA sequences in the training data and using their 3D coordinates as a starting point.

After submitting, I recorded the score:

kaggle-wandb-sync score https://wandb.ai/.../runs/hahu4ygj \
  --tm-score 0.17 \
  --rank 779

All three experiment versions (baseline, improved, template matching) are now tracked in W&B for easy comparison.

→ Notebook: https://www.kaggle.com/code/yasunorim/template-matching-w-b-via-kaggle-wandb-sync

Summary

Command	What it does
run	Full pipeline (push → poll → output → sync)
push	Push notebook to Kaggle
poll	Wait for notebook to finish
output	Download output files
sync	Sync offline runs to W&B
score	Log submission score to W&B run

pip install kaggle-wandb-sync
kaggle-wandb-sync run my-competition/

Kaggle's internet restriction is no longer a blocker for W&B experiment tracking.