A Watermark for Large Language Models

Code

• github.com/jwkirchenbauer/lm-watermarking
  OfficialIn paperpytorch★ 660
• github.com/huggingface/text-generation-inference
  pytorch★ 10,812
• github.com/facebookresearch/three_bricks
  pytorch★ 51
• github.com/chengez/adversarial-paraphrasing
  pytorch★ 37
• github.com/eva-giboulot/watermax
  pytorch★ 8
• github.com/Xieyangxinyu/Unbiased-Watermark-via-Maximal-Coupling
  pytorch★ 3
• github.com/fyyfu/semantic-watermark
  pytorch★ 2
• github.com/BrianPulfer/LMWatermark
  pytorch★ 0

Abstract

Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of "green" tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark. We test the watermark using a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, and discuss robustness and security.

Tasks

Reproductions