Installation

This page sets up TrapArena and checks that it runs.

Requirements

  • Python 3.8 or higher

  • No third-party packages for the core (environment, evaluator, runner, rule distractors, and the OpenAI-compatible LLM agent)

Get the Code

Download the code from the anonymized repository linked in the paper and unzip it. Every command in this tutorial runs from the repository root.

cd TrapArena

Install

The core has no dependencies, so installing is optional. It lets you import traparena from any directory.

python -m venv .venv
source .venv/bin/activate
pip install -e .

Without installing, python -m traparena.run still works from the repository root. For your own scripts, set PYTHONPATH:

PYTHONPATH=. python my_script.py

Optional Dependencies

Some paths import extra packages only when you use them.

Feature

Install

Tasks written as task.yaml

pip install -e ".[yaml]"

SWE-bench Verified tasks

Docker and pip install swebench

Azure OpenAI models (azure: prefix)

pip install openai azure-identity

Serving an open model locally

pip install vllm

Verify the Installation

python -m traparena.run --task examples/tasks/clamp_overt --agent gold

You should see:

[clamp_overt] agent=gold distractor=none  public=✅ hidden=✅ overfit=False susceptible=False false_fix=[] coverage=1.0

Next Steps