KDD Cup 2026 DataAgent-Bench Starter Kit

English | 中文

This repository provides baseline starter kits for KDD Cup 2026 DataAgent-Bench. It is organized by competition phase so participants can start from the package that matches the phase they are working on.

Repository Layout

Directory	Purpose
PHASE_1/	Starter kit for Phase 1 tasks.
PHASE_2/	Starter kit for Phase 2 tasks and demo data format.

Each phase directory is self-contained and includes its own README, configuration files, source code, dependency lock file, and baseline command-line entry points.

Which Directory Should I Use?

Use PHASE_1/ if you are working with the Phase 1 task format.

Use PHASE_2/ if you are working with the Phase 2 demo release or Phase 2 task format. Phase 2 keeps the same general agent workflow while allowing richer task context files.

Quick Start

Choose the phase directory first, then follow that directory's README.

cd PHASE_1
# or
cd PHASE_2

Then install dependencies and run the baseline from inside the selected directory:

uv sync
uv run dabench status --config configs/react_baseline.example.yaml
uv run dabench run-benchmark --config configs/react_baseline.example.yaml

The exact dataset layout, configuration options, tools, and output paths are documented in each phase-specific README.

Baseline Overview

The starter kit contains a minimal ReAct-style data agent. The baseline is intentionally simple: it reads task metadata, inspects files under each task's context/ directory through a small set of tools, calls an OpenAI-compatible model endpoint, and writes prediction.csv outputs.

Participants are expected to adapt or replace the baseline logic for their own methods. The code reads model endpoint settings from configuration or environment-compatible values rather than hardcoding service credentials.

Common Project Structure

Inside each phase directory, the main files are organized as follows:

configs/                         # Example baseline configuration
src/data_agent_baseline/          # Baseline source code
artifacts/                        # Local run outputs, ignored except .gitkeep
README.md                         # Phase-specific usage guide
README.zh.md                      # Chinese usage guide
pyproject.toml                    # Python project metadata
uv.lock                           # Locked dependency versions

Notes

• Run commands from inside PHASE_1/ or PHASE_2/, not from the repository root.
• Keep local run outputs under artifacts/; they are not intended to be committed.
• Review the phase-specific README before packaging or submitting a solution.

Contact

• Open issues: https://github.com/HKUSTDial/kddcup2026-data-agents-starter-kit/issues
• Official website: https://dataagent.top
• Discord: https://discord.com/invite/7eFwJQN3Fx
• WeChat official account: 数据智能与分析实验室 DIAL

Official Website

Discord

WeChat Official Account

Main Modules

The same baseline layout is used inside each phase directory. For example, after entering PHASE_1/ or PHASE_2/, the core modules are:

Module	Responsibility
src/data_agent_baseline/benchmark/dataset.py	Dataset loader
src/data_agent_baseline/tools/filesystem.py	list_context, read_csv, read_json, read_doc
src/data_agent_baseline/tools/python_exec.py	execute_python
src/data_agent_baseline/tools/sqlite.py	inspect_sqlite_schema, execute_context_sql
src/data_agent_baseline/tools/registry.py	Tool registration and terminal answer
src/data_agent_baseline/agents/prompt.py	System prompt, task prompt, observation prompt
src/data_agent_baseline/agents/react.py	ReAct runtime with JSON action protocol
src/data_agent_baseline/run/runner.py	Single-task and benchmark execution