---

name: competition-data-science description: "Prepare datasets and operate RD-Agent data-science/Kaggle workflows with reproducible debug, validation, submission, and grading artifacts." metadata: disco-role: operating parent-skill: rd-agent license: MIT disable-model-invocation: true

RD-Agent competition and data science

Use this sub-skill for rdagent data_science, Kaggle-style tasks, custom datasets, data preparation, generated modeling code, local evaluation, and submission artifacts.

Input contract

A custom task should be organized around a task directory and a source_data root. The repository's example documents the following shape:

source_data/<task>/prepare.py
<task>/<task>/description.md
<task>/<task>/sample.py        # optional debug sampler
<task>/eval/grade.py           # optional test grader
<task>/eval/valid.py           # optional submission validator

The task description should state the objective, data fields, split/use restrictions, modeling constraints, metric, and submission format. Run check_dataset_layout.py before asking an agent to write modeling code.

Safe workflow

1. Inspect rdagent data_science --help and the scenario config.
2. Run prepare.py or the dataset preparation step in a disposable output location; verify row counts, columns, labels, and train/test separation.
3. Create a tiny debug sample using the task's sample.py when present. The default sampler in rdagent.scenarios.data_science.debug.data is a fallback, not a guarantee of semantic correctness.
4. Launch the smallest agent loop that exercises code generation and evaluation.
5. Validate the generated submission with valid.py when available, then grade with grade.py or the configured evaluator.
6. Preserve the description, resolved paths, generated source, submission, validation output, metric, and seed.

Competition hygiene

• Do not use test labels, competition metadata, or future information in feature engineering.
• Check that local validation matches the competition's target and that a generated submission has the exact required columns/order.
• Treat external Kaggle credentials and downloads as prerequisites; do not include them in a reusable skill or test fixture.
• A successful code-generation iteration is not a score. Report the evaluator command and the data split that produced the score.
• For time-series or patient data, preserve group/entity boundaries and temporal ordering while sampling.

Read dataset-contract.md for required files and artifact semantics. For Qlib-backed finance tasks use quant-finance instead.