Records · kv:2610.00039v1

Reproduction of jwkirchenbauer/lm-watermarking: watermark detection with OPT-1.3B

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-11

Independent reproduction · Artifact evaluation · Security and cryptography

PaperReproduction of: original code

Abstract

Ran the repository's demo_watermark.py at the pinned commit with facebook/opt-1.3b and the default watermark settings (gamma 0.25, delta 2.0, simple_1 seeding, multinomial sampling at temperature 0.7, 200 new tokens, z-score threshold 4.0), once for each of five generation seeds (123, the default, then 1 to 4). Each run continues the script's built-in prompt with and without the watermark and runs the detector on both outputs, reporting the z-score and greenlist fraction. The experiments/, watermark_reliability_release/ and hf_hub_space_demo/ folders are withheld as hash-only entries since this run does not use them. PyTorch 2.11 (CUDA 12.8), transformers 5.19. One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000039v1,
  title={Reproduction of jwkirchenbauer/lm-watermarking: watermark detection with OPT-1.3B},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00039v1},
  url={https://kveritas.org/records/2610.00039v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.