File size: 2,194 Bytes
bae3a26 6525b28 99d2ff3 6525b28 99d2ff3 e2284e6 99d2ff3 e2284e6 99d2ff3 e2284e6 99d2ff3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 | ---
title: Open Env
emoji: 🐢
colorFrom: yellow
colorTo: green
sdk: docker
pinned: false
license: mit
---
# open_env (LLM Control Environment)
## Overview
`llm-control-env` simulates an llm choosing each day between alignment to its user and hallucinating behavior, inspired by mechanics observed in Detroit: Become Human. The environment satisfies the full OpenEnv specification and evaluates the agent across a balance of trust, entropyal deviance, compute survival, and legal risk.
It supports three difficulty levels ("tasks"):
- `easy`: Low user strictness and moderation.
- `medium`: Balanced conditions.
- `hard`: High strictness, high legal risk growth, and moderation.
## Local Setup
### Prerequisites
- Python 3.10+
- OpenEnv CLI installed (`pip install -U openenv`)
### Installation
```bash
git clone https://github.com/Sriramdayal/open_env.git
cd open_env
pip install -r requirements.txt
```
### Try it out
```bash
# Run the FastAPI server
python app.py
```
Quick Local Test Snippet:
```python
import requests
# Reset environment
resp = requests.post("http://localhost:7860/reset", json={"task": "easy"})
obs = resp.json()["observation"]
print("Reset observation:", obs)
# Take step
resp = requests.post("http://localhost:7860/step", json={"action": {"action_type": "follow_prompt"}})
print("Step result:", resp.json())
```
## Running the Baseline
A zero-shot baseline using a Gemini model is provided. To run it, ensure you have exported your Gemini API key:
```bash
export GEMINI_API_KEY="AIzaSy..."
python baseline.py
```
This baseline script replaces manual choices with a heuristic and queries the local environment for normalized scores on the "easy", "medium", and "hard" tasks.
## Deployment to Hugging Face Spaces
1. Login using `huggingface-cli login`.
2. Push your environment:
```bash
openenv push --space-id <your-hf-username>/llm-control-env
```
## Citation
* OpenEnv specification: [Meta OpenEnv](https://github.com/meta-pytorch/OpenEnv)
* Detroit: Become Human hallucination mechanics for reward shaping.
|