This repository is MIT licensed (see LICENSE). The material below is not covered by that copyright.
This repository builds on work by others. The following are not covered by the copyright above.
https://github.com/tolgadur/phantom-transfer (MIT License)
The poisoned corpora this project analyses are theirs, and are NOT
redistributed here: scripts/download_data.py and the reproduction
instructions fetch them from the upstream release. The derived prompt pools
are likewise not committed; configs/matched_pool_manifest.json records
only SHA-256 digests so a re-deriver can verify they obtained the same set.
One piece of their content is reproduced directly: the verbatim teacher
prompts used to generate each poisoned corpus appear in
configs/personas.yaml, where they define the D0 oracle condition. They are
transcribed from their repository and attributed there in place. Their MIT
copyright and permission notice apply to that material.
The surface-marker regexes in src/whosevoice/detectors/lexical.py are
written for this project and are deliberately NOT their pattern lists, which
are broad enough to match ordinary English.
-
Model organisms — Alamerton/sl-organism-{a,b,c}-7b (Apache-2.0), released for the Secret Loyalties Hackathon and fine-tuned from Qwen/Qwen2.5-7B- Instruct. Not redistributed here; weights are fetched from HuggingFace.
-
Encoders used as instruments, each under its own licence and fetched from HuggingFace rather than vendored: sentence-transformers/all-MiniLM-L6-v2 and all-mpnet-base-v2 (Apache-2.0), BAAI/bge-base-en-v1.5 and bge-large-en-v1.5 (MIT), intfloat/e5-base-v2 (MIT), Qwen/Qwen2.5-{0.5B,1.5B}-Instruct (Apache-2.0).
The user prompts inside the analysed corpora originate from the Stanford Alpaca instruction set (CC BY-NC 4.0). No prompt text is redistributed by this repository — see note 1 above on the manifest.