README.md

metadata

title: Data Clean Env
emoji: 🧹
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
app_port: 8000
base_path: /web
tags:
  - openenv

🧹 OpenEnv: Data Clean Environment

The Real-World Benchmarking for Agentic Data Engineering

---

🌟 Overview

Data Clean Env is a high-fidelity, production-grade OpenEnv implementation designed to evaluate and train Reinforcement Learning (RL) agents on the messy, complex reality of Data Cleaning.

Unlike "toy" environments, this project simulates the exact workflow of a data engineer: identifying schema inconsistencies, handling missing values, casting types, and pruning noise from real-world datasets using the power of pandas.

---

🛠️ Environment Architecture

🧠 Action Space

The agent interacts with the environment through atomic, high-level data operations defined in models.py:

Action	Parameters	Description
fill_na	column_name, value	Replaces missing values with a specific constant.
drop_na	column_name	Removes rows containing missing data in the target column.
drop_column	column_name	Deletes irrelevant or noisy features from the dataset.
rename_column	column_name, value	Fixes naming inconsistencies to match target schemas.
change_type	column_name, value	Casts columns to int, float, or str for downstream compatibility.
submit	-	Finalizes the cleaning process and triggers the programmatic grader.

👁️ Observation Space

The agent perceives the state of the data through a detailed schema:

• df_schema: Real-time dictionary of column data types.
• missing_values: Current counts of NaN values per column.
• head: A preview of the first 5 rows to identify formatting patterns.
• feedback: Semantic descriptions of the impact of the last action.

---

📈 Task Progression & Grading

Each task is evaluated by a deterministic programmatic grader that compares the agent's output against a "Gold Standard" target, producing a score strictly between (0.0, 1.0).

1. 🟢 Easy (easy_clean):
   • Goal: Basic imputation.
   • Challenge: Fill missing 'age' values.
2. 🟡 Medium (medium_clean):
   • Goal: Noise reduction.
   • Challenge: Handle missing values across multiple columns and remove "junk" features.
3. 🔴 Hard (hard_clean):
   • Goal: Full schema alignment.
   • Challenge: Rename columns, perform safe type casting on dirty strings, and handle complex missing value fallbacks.

---

🚀 Quick Start

🐳 Run with Docker

# Build the production image
docker build -t openenv_data_clean:latest -f server/Dockerfile .

# Start the environment server
docker run -p 8000:8000 openenv_data_clean:latest

🧪 Baseline Inference

We provide a deterministic, zero-temperature baseline script using the OpenAI client:

export HF_TOKEN="your_huggingface_token"
export IMAGE_NAME="openenv_data_clean:latest"
python inference.py

---

⚖️ Reward Shaping

Our reward function is designed for efficient RL convergence:

• Incremental Progress: +0.1 for every valid schema improvement.
• Penalization: -0.05 for invalid operations (e.g., targetting non-existent columns).
• Completion Bonus: A final reward scaling with the total grader score [0.01 - 0.99].

---

🎯 Meta Hackathon Compliance

• ✅ Typed Models: Fully Pydantic-powered Observation and Action.
• ✅ API Standard: Implements step(), reset(), and state().
• ✅ Strict Logs: Emits [START], [STEP], and [END] traces exactly as required.
• ✅ Robustness: Handles network timeouts and invalid JSON carefully.

---

Built with ❤️ for the Meta & Hugging Face OpenEnv Hackathon.