BlueGlass is an open-source framework for interpretability and analysis of vision-language and vision-only models. It empowers researchers and practitioners to understand what deep models see and focus on by offering tools to extract, probe, and build ad hoc XAI models for internal representations.
The framework includes a custom infrastructure and lightweight modifications to third-party model libraries, enabling seamless feature extraction from the models. These extracted features are compiled into structured datasets called BlueLens, laying the foundation for advanced Explainable AI research and experimentation.
-
Benchmarking
Seamless support for benchmarking different models with various datasets and evaluators. -
-
Effortlessly generate feature datasets for training ad-hoc models by capturing model activations using a custom interceptor manager - Feature Recorder
-
-
🧪 Ad-hoc Interpretability Models
Leverage the BlueLens dataset to train models like:- Linear Probes
- Sparse Autoencoders (SAEs)
- Other direct interpretability techniques
All optimized for large batch sizes and efficient performance.
-
🧰 Generate Custom BlueLens Dataset
Use the built-in tools to generate your own BlueLens dataset from any model. -
🔁 Feature Patching & Re-Integration
Ad-hoc models can be patched back into the base model using the Feature Patcher, enabling:- Validation of interpretability methods
- Performance benchmarking of modified models
-
🚀 Streamlined Runner
A simplified, plug-and-play design built around modular runners and interfaces for training, validating, and evaluating ad-hoc models. -
⚙️ Scalable Training
- Native support for distributed training
- Central control over precision (e.g., FP16, BF16)
-
🔍 Interpretability & Analysis
Built for both vision-language and vision-only models with a focus on clarity, transparency, and scalability.
- 📚 Tutorials and example notebooks
Coming soon: Installation guide, usage examples, and API reference.
We welcome contributions! Feel free to open issues or submit pull requests to help expand the BlueGlass ecosystem.