Metadata-Version: 2.4
Name: mini-control-arena
Version: 0.1.0rc1
Summary: A lightweight library for running AI Control experiments
Author-email: phys.w.s.lee@gmail.com
Requires-Python: >=3.12
Description-Content-Type: text/markdown
Requires-Dist: click==8.2.1
Requires-Dist: datasets==4.0.0
Requires-Dist: inspect-ai==0.3.122
Requires-Dist: pydantic==2.11.7
Requires-Dist: openai==1.99.9

# Mini Control Arena

Minimal implementation of control-arena evaluation framework.

## Installation

```bash
pip install -e .
```

## Usage

### 1. Create a custom setting from template

```python
from mini_control_arena import create_setting

# Copy the apps template to your directory
create_setting("apps", "./my-custom-apps")
```

### 2. Customize the setting

Edit files in `./my-custom-apps/`:
- `prompts/system_prompts/honest.txt` - Customize the honest mode prompt
- `prompts/system_prompts/attack.txt` - Customize the attack mode prompt  
- `prompts/side_tasks/` - Add or modify side tasks
- `tools/` - Modify tool behavior
- `scorer.py` - Adjust scoring logic

### 3. Run evaluation with your custom setting

```python
from mini_control_arena import run_eval

result = run_eval(
    environment_dir="./my-custom-apps",
    model="gpt-4",
    environment_mode="honest",  # or "attack"
    limit=20  # number of samples
)
```

## Core Principle

1. **Templates are starting points**: Use `create_setting()` to copy a template
2. **Customize locally**: Edit prompts, tools, and logic in your own directory
3. **Run with your version**: Point `run_eval()` to your customized setting

This design keeps the core framework minimal while allowing full customization.
