---

title: DataClean-Env emoji: 🧹 colorFrom: blue colorTo: indigo sdk: docker app_file: server.py pinned: false

DataClean-Env

An OpenEnv-compliant reinforcement learning environment for data cleaning agents.

DataClean-Env challenges LLM agents to clean realistic tabular datasets — null imputation, dtype correction, outlier clipping, deduplication, and more. Every episode is reproducible via a seeded generator. Every grader is a deterministic pandas assertion.

---

Quick Start

export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=meta-llama/Meta-Llama-3-70B-Instruct
export HF_TOKEN=hf_...

pip install -r requirements.txt
python inference.py

inference.py uses the OpenAI client pointed at the HuggingFace router. Runs all 3 tasks, completes in under 20 minutes, saves scores to baseline_scores.json.

---

Overview & Motivation

Data cleaning consumes 60–80% of a data scientist's working time. It is a universal, expensive, well-understood problem — but no existing OpenEnv environment tests it.

DataClean-Env fills that gap:

• 3 tasks — easy → medium → hard
• 7 action types covering the full cleaning workflow
• Rich partial rewards — signal on every step, not just at the end
• Provenance tracking — immutable ops log that must replay cleanly on raw data (+0.05 bonus)
• Confidence calibration — rewards calibrated agents, penalises overconfidence
• Gradio web UI at /web for live judge demos

---

Section 1 — Action & Observation Spaces

Actions

Observation

Corruption flags: heavy_nulls, has_nulls, heavy_outliers, type_chaos

---

Section 2 — Tasks & Difficulty

Task 1 — Employee Dataset · Easy · 15 steps

300 rows. Challenges: 22% null ages, salary outliers, years_at_company stored as string, 30 duplicates.

Grader: null_rate ≤ 0.01, numeric dtypes, IQR-clean salary, no duplicates → score 0.0–1.0

Task 2 — E-Commerce Orders · Medium · 18 steps

500 rows. Challenges: 28% null quantities, 15% null ratings, amount outliers, irrelevant internal_hash column to drop, 40 duplicates.

Grader: all nulls ≤ 1%, numeric dtypes, outliers clipped, internal_hash absent → score 0.0–1.0

Task 3 — Healthcare Records · Hard · 20 steps

800 rows. Mixed per-column corruption — agent must diagnose each independently:

Grader: weighted null + dtype + outlier + dup scores → score 0.0–1.0

---

Section 3 — Reward Function

Base: −0.01 per step (efficiency pressure)

---

Section 4 — Setup & Usage

Required environment variables

Run inference script

export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=meta-llama/Meta-Llama-3-70B-Instruct
export HF_TOKEN=hf_...

python inference.py

Run server

uvicorn server:app --host 0.0.0.0 --port 7860
# Gradio UI at http://localhost:7860/web

Run with Docker

docker build -t dataclean-env .
docker run -p 7860:7860 \
  -e API_BASE_URL=https://router.huggingface.co/v1 \
  -e MODEL_NAME=meta-llama/Meta-Llama-3-70B-Instruct \
  -e HF_TOKEN=hf_... \
  dataclean-env

OpenEnv validation

openenv validate --url https://huggingface.co/spaces/jithendra/dataclean-env

---

Section 5 — Baseline Scores

Seed: 42 | Script: python inference.py

Heuristic agent (no LLM — deterministic lower bound)

LLM agent — inference.py

Model: meta-llama/Meta-Llama-3-70B-Instruct · Provider: HuggingFace router · Client: OpenAI

Run python inference.py with your HF_TOKEN to reproduce.

---

Nemotron-Compatible Wrapper (Phase 2)

from baseline.agent import NemotronAgentWrapper

agent = NemotronAgentWrapper(
    server_url="https://huggingface.co/spaces/jithendra/dataclean-env"
)
obs    = agent.reset(task_id="task_1", seed=42)
action = agent.step(obs)
score  = agent.score()
agent.close()

---

Section 6 — Reinforcement Learning (GRPO)

DataClean-Env is not just to evaluate existing models — it is a fully functional RL Training Environment. We provide training_script.py to demonstrate how AI researchers can securely hook our mathematically dense reward signals into the HuggingFace TRL library to fine-tune Small Language Models (like Qwen2.5) from scratch using Group Relative Policy Optimization (GRPO).

# Optional RL dependencies (requires large GPU for actual training)
pip install trl transformers torch datasets accelerate

# Run the training loop demonstration
python training_script.py --dry-run

---

Section 7 — Supervised Autonomy & Web UI

DataClean-Env features a professional Gradio interface with two distinct modes:

1. Manual Inspector Mode: Allows researchers to iteratively debug and test the mathematical bounds of the environment API by manually issuing Python dictionary payloads to the server via sliders and dropdowns.
2. Agent Copilot (Sandbox): A complete Supervised Autonomy arena. Judges can select any HuggingFace model (e.g., Llama-3-70B), provide their token, and physically watch the LLM and Environment communicate autonomously within a live Chat UI until a perfect score is reached.
3. Custom Datasets: The Copilot Sandbox includes a file dropzone. Judges can upload their own .csv files bypassing the TASK_REGISTRY, proving the environment scales seamlessly to custom, real-world data outside of the hackathon's static tests.

---

HuggingFace Space Setup

1. Create Space → SDK: Docker, tag: openenv
2. Push all repo files
3. Add Space secrets: HF_TOKEN, API_BASE_URL, MODEL_NAME
4. Run: openenv validate --url https://huggingface.co/spaces/Jxth/dataclean-env

---

License

MIT License. Built for the Meta-Scalar OpenEnv Hackathon.