Kaggle

Run the same YAML on Kaggle GPUs. TRLoom can:

1. Detect that it is already inside a Kaggle kernel and train inline
2. Generate a notebook (+ kernel-metadata.json) for upload
3. Push that notebook via the Kaggle API (kaggle.push: true)

Outputs default to /kaggle/working/outputs.

Setup

pip install "trloom[kaggle]"
# Configure API credentials (for push):
#   ~/.kaggle/kaggle.json  or  KAGGLE_USERNAME + KAGGLE_KEY

In a Kaggle notebook session, also enable a GPU accelerator when you need CUDA.

YAML settings

kaggle:
  enabled: true
  mode: auto              # auto | inline | notebook
  accelerator: gpu        # gpu | tpu | none
  push: false
  username: null          # or set KAGGLE_USERNAME
  kernel_slug: null       # defaults to trloom-<config-stem>
  title: TRLoom Fine-Tune
  is_private: true
  enable_internet: true
  dataset_sources: []
  competition_sources: []
  kernel_sources: []
  pip_packages: []
  install_source: pypi    # pypi | git
  git_url: git+https://github.com/saqlain2204/trloom.git
  output_dir: /kaggle/working/outputs
  notebook_path: null
  download_dir: null
  wait: true
  timeout_seconds: 32400

Field	Purpose
enabled	Prefer the Kaggle backend
mode	auto picks inline vs notebook; inline runs now; notebook writes an .ipynb
accelerator	Written into kernel metadata (gpu / tpu / none)
push	Push the generated notebook with the Kaggle API
install_source	How TRLoom is installed inside the notebook
output_dir	Training output path inside the kernel

modal.enabled and kaggle.enabled are mutually exclusive.

User code and formatters

Local callables referenced from YAML (map_fn, formatting_func, rewards, …) are bundled into the notebook the same way they are for Modal, so remote Kaggle runs can import them.

dataset:
  map_fn: formatters:instruction_to_text

user_code:
  - ./formatters.py

kaggle:
  enabled: true
  mode: notebook

CLI

# Force Kaggle backend
trloom run examples/sft_kaggle.yaml --kaggle

# Write notebook + kernel-metadata.json
trloom kaggle-notebook examples/sft_kaggle.yaml -o ./kaggle_job/trloom.ipynb

Python API

from trloom.kaggle_support import run_on_kaggle, write_kaggle_notebook

write_kaggle_notebook("examples/sft_kaggle.yaml", destination="out/job.ipynb")
run_on_kaggle("examples/sft_kaggle.yaml")

Inside a Kaggle kernel, mode: auto (or inline) disables nested remote dispatch and redirects training.output_dir to kaggle.output_dir.