This is repository for reproducing GAR on Reasoning Intensive IR

Installation

Prerequisites

• Python 3.10+ (Python 3.12 recommended)
• CUDA-compatible GPU (for local model inference)
• Java 11+ (for PyTerrier)

Setup

1. Create a virtual environment:

   python3 -m venv venv
   source venv/bin/activate

2. Install dependencies:

   pip install -r requirements.txt
   # Additional PyTerrier plugins
   pip install git+https://github.com/terrierteam/pyterrier_pisa.git
   pip install --upgrade git+https://github.com/emory-irlab/pyterrier_genrank.git
   pip install --upgrade git+https://github.com/terrierteam/pyterrier_t5.git

Preprocessing: Index and Graph Creation

Before running experiments, you may need to create an index and a corpus graph.

1. Create Index

Use create_index.py to index the dataset.

python3 create_index.py --task <TASK_NAME> 

Arguments:

• --task: The specific task from the BRIGHT benchmark (default: biology).

2. Create Graph

Use create_graph.py to create a corpus graph for adaptive reranking.

python3 create_graph.py --task <TASK_NAME> --k <K>

Arguments:

• --task: The specific task (default: biology).
• --k: Number of neighbors in the graph (default: 16).

Usage

The main entry point for running experiments is run.py.

Running an Experiment

To run a reranking experiment on a specific task within the BRIGHT benchmark:

python3 run.py --task <TASK_NAME> --model_name <MODEL> --budget <TOP_K>

Examples

Run Rank1-7B on the Robotics task with a budget of 100 documents:

python3 run.py --task robotics --model_name rank1-7b --budget 100

Run TFRank-8B on the Sustainable Living task:

python3 run.py --task sustainable-living --model_name tfrank-8b --budget 50

Budget 50 results

Due to limited space, we added budget c=50 results per subtask in budget_50.md