🤖 AutoData Agent – Autonomous Hugging Face AI for Data Cleaning & Modeling

An autonomous data science agent that cleans, explores and models your data intelligently 🧠

Built with Hugging Face SmolAgents, Ollama, and Scikit-Learn

---

🧠 Project Overview

AutoData Agent is a modular and autonomous data assistant designed to:

• 🧹 Inspect raw datasets (missing values, data types, anomalies)
• 🧼 Clean and preprocess data (imputation, encoding, scaling)
• 📊 Visualize key statistical properties (distributions, correlations, boxplots)
• 🤖 Train machine learning models automatically (classification or regression)
• ⚙️ Optimize the model using GridSearchCV to achieve best performance

This project demonstrates how a Hugging Face agent can orchestrate an end-to-end Data Science pipeline, making smart decisions and reasoning about the dataset structure.

---

🧩 Architecture

datacleaner-agent/
├── app.py                 # Main entry point
├── agent_logic.py         # Builds the Hugging Face Agent (model + tools)
├── tools/
│   ├── inspect.py         # InspectTool: automatic EDA (plots + summary)
│   ├── cleaning.py        # CleaningTool: data cleaning & encoding
│   └── train.py           # TrainTool: automatic ML training + evaluation
├── test_tools.py          # Local tests for each tool
├── requirements.txt
└── README.md

---

🚀 How It Works

🔹 Step 1 — InspectTool

Performs an Exploratory Data Analysis (EDA):

• Displays dataset info (shape, dtypes, missing values)
• Generates histograms, boxplots, and correlation heatmaps
• Detects data imbalances and null distributions

Example output:

• df.info(), df.describe()
• Automatic visualizations for numeric variables
• Summary of missing data

---

🔹 Step 2 — CleaningTool

Cleans and prepares the dataset:

• Removes duplicates
• Handles missing values (median or mode)
• Encodes categorical variables (LabelEncoder or OneHotEncoder)
• Scales numerical columns (StandardScaler)
• Detects and drops low-variance features

Goal: produce a dataset ready for model training.

---

🔹 Step 3 — TrainTool

Automatically trains a model based on target variable type:

• Detects whether it’s classification or regression

• Chooses the appropriate RandomForest model

• Runs GridSearchCV to optimize hyperparameters

• Splits data into train/test (80/20)

• Displays key metrics:

  • Classification: Accuracy, Precision, Recall, F1
  • Regression: RMSE, R²

---

⚙️ Installation

Prerequisites

• Python ≥ 3.11
• Virtual environment recommended

git clone https://github.com/MeidiLprog/datacleaner-agent.git
cd datacleaner-agent
pip install -r requirements.txt

---

🧭 Usage

▶️ Run the agent

python app.py

The agent will:

1. Load the Titanic dataset (default)
2. Inspect and clean it automatically
3. Train a predictive model on Survived
4. Output key metrics and model summary

Expected Output:

Dataset successfully loaded ! (891, 12)
Agent ready !
Inspecting dataset...
Cleaning done...
GridSearch training...
Accuracy: 0.84
F1 Score: 0.81

---

☁️ Supported Execution Modes

🟡 Hugging Face Cloud (Recommended)

Runs the reasoning model via Hugging Face Inference API.

Set your token:

export HUGGINGFACEHUB_API_TOKEN="hf_xxxxxxxxxxxxxxxxxxxxxxxxx"

Then in agent_logic.py:

model = LiteLLMModel(model_id="huggingface/mistralai/Mistral-7B-Instruct-v0.2")

🔵 Local (Offline) – Ollama

If you prefer running locally:

1. Install Ollama

2. Pull the model:

   ollama pull qwen2:1.5b

3. Replace model in agent_logic.py:

   model = LiteLLMModel(model_id="ollama/qwen2:1.5b")

---

🧰 Technologies Used

Stack	Purpose
🤗 Hugging Face SmolAgents	Agent orchestration
🔮 LiteLLM / HfApiModel	LLM reasoning
🧹 Pandas / Numpy	Data wrangling
📊 Matplotlib / Seaborn	Data visualization
⚙️ Scikit-Learn	Model training & GridSearch
💻 Ollama	Local LLM inference (offline mode)

---

💡 Example Screenshots

Visualization	Description
	Automatic data histograms & boxplots
	Correlation matrix
	Model training output

---

🧑‍💻 Author

Lefki Meidi 🎓 Data Science & Machine Learning Engineer 💬 LinkedIn • GitHub • HuggingFace

---

🌟 Project Highlights

• Built entirely from scratch in less than 24h
• Modular architecture (plug & play tools)
• Hugging Face AI agent integrated locally and via cloud API
• Fully autonomous workflow: from raw data → cleaned dataset → trained model
• Ideal for data preprocessing automation or teaching agent reasoning

---

❤️ Acknowledgements

Special thanks to Hugging Face for the SmolAgents framework, and the open-source community for making AI accessible.

“Why spend hours cleaning data when your agent can do it for you?”