QSBC — Quantized Semantic Bottleneck Classifier
Interpretable sentence classification via a semantic proposition bottleneck
About
QSBC replaces the direct text-to-label mapping of an uninterpretable transformer with a two-stage path: concept extraction, then a strictly linear probe, so concept attribution to class logits is analytically exact. This repo builds the data layer and experiments: a Dockerized PostgreSQL corpus of the parsed C3PA privacy policies matching the reference split, a chain-of-thought `cot` agent that generates atomic generalized rationales per sentence-label pair, the Phase 2 thoughts/ideas alphabet (statements clustered into canonical concepts), a FastAPI + Bootstrap SPA explorer, and Colab experiment notebooks (exp1-4) benchmarking the pipeline with a linear-probe ceiling reported at +9.8pt.
Summary
Infrastructure and experiments for QSBC, an interpretable sentence-classification architecture that inserts a human-meaningful semantic proposition bottleneck between input text and linear-probe prediction, applied to the CPRA subset of the C3PA privacy-policy dataset.
Tech
- python
- fastapi
- uvicorn
- postgresql
- docker
- scikit-learn
- sentence-transformers
- jupyter
Tags
- classification
- interpretability
- nlp
- privacy-policies
- semantic-bottleneck