Ash1104/data-cleaning-env

Data Cleaning Agent Environment

Motivation and Description

This OpenEnv simulation models an exceptionally common, real-world task natively critical to nearly every data engineering and MLOps pipeline: Dataset Triage & Cleaning. Agents interact with simulated user datasets corrupted natively with human-like errors (null fields, duplicate injections, malformed target variables, and unpredictable datetime formats). Agents must execute sequences of programmatic transformations culminating in a verified SQL abstraction that perfectly filters anomalies.

Spaces

Observation Space

A structured representation mapping the environment boundaries:

• —table_sample: List representing the first 5 active rows as dictionaries.
• —columns: Metadata detailing the schema (name and dtype lists).
• —step: Current progression step.
• —max_steps: Total allowed steps for the trajectory.

Action Space

A dynamic schema expecting BaseModel actions matching:

• —type: "label_issue": Identify a missing structure (requires row, col, issue_type).
• —type: "fix_cell": Submit a normalized string/numeric value (requires row, col, value).
• —type: "drop_row": Remove explicitly duplicated records natively.
• —type: "submit_sql": Build arbitrary DuckDB SQL against raw_data mapped to a valid result.
• —type: "finish": Manual early-termination command natively capping the score.

Tasks and Difficulty Ranges

1. 1.Easy (20 max steps): 4 injected issues across 100 rows natively testing LLM extraction skills.
2. 2.Medium (30 max steps): 8 natively scattered issues demanding consistent variable formatting mapping.
3. 3.Hard (40 max steps): 12 layered complex issues heavily tested against stringent DuckDB constraint checks in validation (id bounds, NULL checks, REGEX matches).

Setup & Validation Instructions

1. 1.Build and Run the Engine:

   docker build -t dq-env .
   docker run -p 7860:7860 dq-env

1. 1.Verify Spec Strictness:

   openenv validate openenv.yaml --endpoint http://localhost:7860

1. 1.Run Local Inference Baseline: Map your endpoint context naturally via OpenEnv conventions:

   python inference.py

Immediate Baseline Results

The active inference.py evaluates across all trajectories pulling baseline marks matching 0.98 overall accuracy globally inside of 1 second execution windows. Trajectories execute deterministically with seeded generation mechanics to optimize inference testing stability natively matching OpenEnv requirements.