Thinxai

A CLI tool for explaining deep learning models. Provide it with a trained model + training dataset with target and it generates a PDF report telling what drives predictions and why.

Install

pip install thinxai-II

With extras:

pip install thinxai[tensorflow]  
pip install thinxai[gpu]          
pip install thinxai[all]           

Quick Start

1. Place your model file and CSV dataset in the same folder
2. Run:

thinxai

Or:

python -m thinxai

3. Select your model and dataset from the interactive menu
4. Find the PDF report in the same folder

Supported Inputs

Deep Learning Models: PyTorch (.pth, .pt, .pkl), TensorFlow/Keras (.h5, .hdf5, .keras, SavedModel), Transformers (BERT, GPT, etc.), ONNX (.onnx)

Also Supported: Scikit-learn (.pkl, .joblib)

Data: CSV, tab-delimited, pipe-delimited

What It Does

Thinxai runs multiple explainability methods on your model and combines the results into a single ranked list. It then generates plain English explanations for the top features and packages everything into a PDF report.

Analysis Pipeline

• Permutation Importance: shuffles each feature and measures accuracy drop
• SHAP Values: game-theoretic attribution per prediction
• Integrated Gradients: gradient-based attribution along the input path
• Statistical Tests: mutual information and F-scores
• Built-in Importance: model native scores when available

These are merged into a consensus score. Features that rank highly across multiple methods get higher confidence.

Explanations

Top features are explained by a cascade LLM system:

• Groq API: Primary
• HuggingFace fallback: Secondary
• Rule-based fallback: Always works

Each explanation includes a business insight

PDF Report Contents

• Cover page with model validation metrics and trust indicator
• Feature importance distribution pie chart
• Each feature gets explanation cards with business insights
• Executive summary with the top 3 features
• Technical glossary automatically generated from terms found in the report

Trust Indicators

The tool flags suspicious results:

• Accuracy >99%: warns about possible data leakage
• Accuracy near random: warns model may not have learned patterns
• Raw state_dict models: skips validation, runs statistical analysis only

Requirements

• Python 3.8+