What if you do not know the right taxonomy before reading the data?
A taxonomy that can evolve
Many classification tasks start with a fixed list of labels. This project explores the other direction: an LLM assigns labels to text and can introduce categories as it encounters new material.
A configurable limit keeps the label set manageable. When consolidation is needed, the system can merge similar categories, while fixed classes remain protected from those merges.
Give the merging step context
The implementation can track category frequency and extract keywords to inform consolidation. That makes the question more specific than whether two labels merely sound similar: their use in the dataset can also influence a decision.
The interface is intentionally small: a Python API and a command-line path for CSV processing. Configuration covers category limits, label length, protected classes and the text column to analyse.
Flexible labels still need checking
An evolving taxonomy trades a predefined structure for adaptability. Its usefulness depends on the task, the model and the review of classification and merge decisions. The repository documents the mechanism; it does not establish a universal accuracy result or eliminate the need to inspect the output.
