LOCA: Logical Chain Augmentation for Scientific Corpus Cleaning

Overview

LOCA (Logical Chain Augmentation) is a novel framework for automatically cleaning scientific corpora, specifically designed to address the high error rates commonly found in scientific question-answering (QA) datasets.

Installation

Prerequisites

• Python 3.12+
• LLM API key
• uv (recommended)

Setup

1. Clone the repository

git clone xxx
cd LOCA

2. Install dependencies

Using uv (recommended):

uv sync

Project Structure

LOCA/
├── src/                         # Source code
│   ├── loca/                    # LOCA implementation
│   │   ├── solver.py            # Main LOCA solver
│   │   ├── results/             # LOCA results
│   │   └── utils/               # Utilities
│   │       ├── augmentation.py  # Augmentation agent
│   │       ├── reviewer.py      # Review agents
│   │       └── secretary.py     # Secretary for summarizing reviews
│   ├── api/                     # LLM API interfaces
│   ...
├── configs/                     # Configuration files
├── problem_set/                 # Test datasets
├── test_results/                # Evaluation results on other methods
└── scripts/                     # Analysis scripts

Usage

Running LOCA

uv run main.py --config PATH_TO_YAML --config-name CONFIG_NAME_IN_YAML

Automated External Consistency Check

./scripts/run_analyze_improved_solutions.sh