README.md

metadata

language:
  - en
pretty_name: PostMark data files
tags:
  - watermarking
  - llm

PostMark data files

Data files needed to run PostMark, from the paper PostMark: A Robust Blackbox Watermark for Large Language Models (EMNLP 2024).

These files were previously hosted on Google Drive. That link no longer works, so they are hosted here instead.

Files

File	Size	Used for
paragram_xxl.pkl	2.0 GB	Paragram word embedding table (a pickled torch tensor). Used by detect.py to compute word presence scores, and loaded by every run. Its vocabulary is paragram_xxl_words.json in the GitHub repo.
filtered_data_100k_unique_250w_sentbound_openai_embs.pkl	231 MB	OpenAI text-embedding-3-large (256-dim) document embeddings, used to build the word embedding table when running with --embedder openai.
filtered_data_100k_unique_250w_sentbound_nomic_embs.pkl	335 MB	nomic-embed-text-v1 document embeddings, used the same way when running with --embedder nomic.

SHA-256 checksums are in SHA256SUMS.

Usage

Download the files into the root of your PostMark checkout:

cd PostMark
hf download yapeichang/PostMark --repo-type dataset --local-dir . \
    --include "*.pkl" "SHA256SUMS"
sha256sum -c SHA256SUMS   # optional: verify the downloads

See the GitHub README for how to run watermarking and detection.

Citation

@misc{chang2024postmark,
      title={PostMark: A Robust Blackbox Watermark for Large Language Models},
      author={Yapei Chang and Kalpesh Krishna and Amir Houmansadr and John Wieting and Mohit Iyyer},
      year={2024},
      eprint={2406.14517},
      archivePrefix={arXiv},
}