PiLLar: Matching for Pivot Table Schema via LLM-guided Monte-Carlo Tree Search

PiLLar is the first framework designed specifically for matching for pivot table schema — aligning pivot tables with standard relational tables using an LLM-driven Monte-Carlo Tree Search with theoretical convergence guarantees.

The framework achieves training-free adaptation, high accuracy with minimal anonymized sample data, and introduces the first benchmark dataset for matching for pivot table schema tasks.

Repository Structure

PiLLar/
├── README.md
├── requirements.txt                  # Python dependencies
├── dataset/                          # All datasets used in experiments
│   ├── adult/
│   │   ├── source.csv                # Pivot table
│   │   ├── target.csv                # Standard table
│   │   ├── column_explanations.json  # Column descriptions
│   │   └── ground_truth.json         # Ground truth mapping
│   └── ...                           # Football, President and Gene datasets
└── src/                              # Source code of the PiLLar framework
    ├── __init__.py
    ├── pillar_globals.py             # Global configuration, shared state, models, caches
    ├── logging_utils.py              # Simple logging helpers
    ├── llm_utils.py                  # LLM client wrappers
    ├── similarity.py                 # Similarity metrics & reward computation
    ├── mcts.py                       # Bounded-stochastic MCTS
    └── main.py                       # Command-line entry point for running PiLLar

Benchmark Datasets

This repository includes the PTBench benchmark introduced in the paper — the first dataset designed for matching for pivot table schema tasks.

Dataset	Type	# Attr. (Pivot → Standard)
Adult	Census	19 → 19
Football	Sports Analytics	23 → 13
President	Evaluation Metrics	12 → 4
Gene	Biological Data	119 → 96

Installation

git clone https://github.com/ZJU-DAILY/PiLLar.git
cd PiLLar
pip install -r requirements.txt

You will need access to an LLM endpoint (Qwen, GPT, Claude, etc.). Specify it via environment variables PiLLar_API_KEY and PiLLar_BASE_URL.

Quick Start

python -m src.main -d football