Apppublic

Sugeerth2/llminterpretabilitypipeline

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

LLM Interpretability Pipeline

A comprehensive, interpretable machine learning pipeline for NLP classification tasks. This pipeline emphasizes transparency and explainability at every stage, from data preprocessing to model predictions.

Live demo

Type any sentence in the box at the top of this Space — you'll see three independent attribution methods (Integrated Gradients, last-layer Attention, SHAP) run side-by-side on DistilBERT fine-tuned for SST-2 sentiment. Watching the three methods agree (or disagree) is a quick read on how robust the model's reasoning is for that input.

Source code: examples/serve_demo.py · API: POST /explain {"text": "...", "top_k": 8}

Features

—Multiple Model Support: BERT, DistilBERT, RoBERTa, and traditional ML models
—Built-in Interpretability:
—SHAP (SHapley Additive exPlanations)
—LIME (Local Interpretable Model-agnostic Explanations)
—Attention Visualization
—Integrated Gradients
—Token Attribution
—End-to-End Pipeline: Data loading → Preprocessing → Training → Evaluation → Interpretation
—Experiment Tracking: Built-in logging and metrics tracking
—Modular Design: Easy to extend and customize

Installation

bash

# Clone the repository
git clone https://github.com/sugeerth/llm-interpretability-pipeline.git
cd llm-interpretability-pipeline

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Quick Start

python

from src.pipeline import InterpretableNLPPipeline

# Initialize pipeline
pipeline = InterpretableNLPPipeline(
    model_name="distilbert-base-uncased",
    num_labels=2,
    interpretability_methods=["shap", "attention", "lime"]
)

# Train on your data
pipeline.fit(train_texts, train_labels)

# Get predictions with explanations
predictions, explanations = pipeline.predict_with_explanations(test_texts)

# Visualize attention patterns
pipeline.visualize_attention(text="This movie was absolutely fantastic!")

# Generate SHAP explanations
pipeline.explain_shap(text="The product quality is terrible.")

Project Structure

llm-interpretability-pipeline/
├── src/
│   ├── models/           # Model architectures
│   ├── interpretability/ # SHAP, LIME, Attention viz
│   ├── data/             # Data loading and preprocessing
│   └── utils/            # Helper functions
├── notebooks/            # Jupyter notebooks with examples
├── configs/              # Configuration files
├── tests/                # Unit tests
└── examples/             # Example scripts

Interpretability Methods

1. SHAP Values

Compute feature importance using Shapley values from game theory.

python

from src.interpretability import SHAPExplainer

explainer = SHAPExplainer(model, tokenizer)
shap_values = explainer.explain(text)
explainer.visualize(shap_values)

2. LIME Explanations

Generate local interpretable explanations.

python

from src.interpretability import LIMEExplainer

explainer = LIMEExplainer(model, tokenizer)
explanation = explainer.explain(text, num_features=10)

3. Attention Visualization

Visualize transformer attention patterns.

python

from src.interpretability import AttentionVisualizer

visualizer = AttentionVisualizer(model, tokenizer)
visualizer.plot_attention_heads(text, layer=11)

4. Integrated Gradients

Attribute predictions to input features using gradients.

python

from src.interpretability import IntegratedGradientsExplainer

explainer = IntegratedGradientsExplainer(model, tokenizer)
attributions = explainer.explain(text)

Configuration

Configure the pipeline using YAML files:

yaml

# configs/default.yaml
model:
  name: "distilbert-base-uncased"
  num_labels: 2
  max_length: 512

training:
  batch_size: 16
  learning_rate: 2e-5
  epochs: 3

interpretability:
  methods: ["shap", "lime", "attention"]
  shap_samples: 100
  lime_num_features: 10

Examples

See the notebooks/ directory for detailed examples:

—01_basic_classification.ipynb - Basic text classification
—02_interpretability_demo.ipynb - All interpretability methods
—03_custom_models.ipynb - Using custom model architectures

License

MIT License

Citation

If you use this pipeline in your research, please cite:

bibtex

@software{llm_interpretability_pipeline,
  title = {LLM Interpretability Pipeline},
  author = {Sugeerth Murugesan},
  year = {2024},
  url = {https://github.com/sugeerth/llm-interpretability-pipeline}
}