Research use only · public proxy corpus
Stylometric detection of LLM-rewritten dark-web documents.
A Colab-faithful lab: word TF–IDF, character n-grams, eleven stylometric features, LinearSVC and logistic regression. Two tasks — human vs rewrite, and the original ten-way topic. No live Tor collection.
Working
Word TF–IDF, character n-grams, LinearSVC — off the UI thread.
RQ1
Can TF–IDF + character n-grams + cheap stylometry separate human text from LLM rewrites of the same documents?
RQ2
Does adding stylometry improve source detection over TF–IDF alone?
RQ3
Does source-aware training hurt or help the original 10-way topic task?
RQ4
Does a detector trained on Model-A rewrites transfer to Model-B rewrites?
What this is — and is not
CoDA (Jin et al., NAACL 2022) is gated for research use. This lab trains on a synthetic proxy with the same ten topic names, masked identifiers, and a controlled paraphrase protocol. It is a measurement bench, not a dump of hidden-service pages.
Underground markets increasingly polish listings with language models. Most published work classifies topic. The second question: can sparse stylometry tell a human-authored document from an LLM paraphrase of the same document, without fine-tuning a transformer?