Mona-2013/openenv-hackathon

OpenEnv Hackathon Project

📌 Overview

This repository contains a real‑world OpenEnv environment with three tasks of increasing difficulty. The environment is containerized with Docker and deployable on Hugging Face Spaces. It follows the OpenEnv specification strictly and passes validation.

🎯 Motivation

The goal is to simulate tasks that humans perform in everyday workflows (email triage, data cleaning, code review) and evaluate LLM agents on structured, reproducible benchmarks.

🧩 Action & Observation Spaces

• —Observation Space: Each task provides structured input (text, dataset rows, or code snippets) via Pydantic models.
• —Action Space: Agents respond with deterministic actions (labels, transformations, or review decisions).
• —Reward Function: Incremental rewards are given for progress toward the objective; penalties apply for invalid or destructive actions.

📝 Tasks

1. Email Triage (Easy)

• —Observation: Raw email text.
• —Action: Classify as spam, important, or other.
• —Reward: +1.0 for correct classification, 0.0 otherwise.

2. Data Cleaning (Medium)

• —Observation: Dataset row with inconsistent formatting.
• —Action: Normalize values, remove duplicates, or clean text.
• —Reward: Incremental reward for each correctly cleaned field.

3. Code Review (Hard)

• —Observation: Code snippet with potential issues.
• —Action: Approve, reject, or comment with feedback.
• —Reward: +1.0 for correct review decision, partial reward for useful comments.

📊 Baseline Results

Baseline scores are reproducible using the provided inference.py.

⚙️ Setup & Usage

Local Run

# Build Docker image
docker build -t openenv .

# Run container
docker run openenv