MarkLLM
THU-BPMMarkLLM is a Python toolkit for embedding watermarks in the text a language model generates and for detecting them. It puts many published methods behind one interface, among them KGW, Unigram, SynthID-Text and the publicly verifiable UPV, and adds visualizations of how each method marks tokens and pipelines that test detection rates, robustness to edits and paraphrase, and text quality. The THU-BPM research group develops it, and it was published as an EMNLP 2024 demo paper.
Tagged self-attested: most of its detectors need the key the generator used, so only the key holder can check a mark.
Openness
5 high confidence- license
- Apache-2.0 (OSI; the LICENSE body is the unmodified Apache 2.0 text)
- source
- public (the toolkit, its algorithm configurations and evaluation pipelines)
- core features withheld
- no — an academic research toolkit with nothing sold beside it, and the tree has no enterprise path
The whole toolkit is Apache-2.0 in one public repository. It is an academic research project with no paid edition, so nothing is held back from the published code.
- https://raw.githubusercontent.com/THU-BPM/MarkLLM/main/LICENSE recorded 2026-11-12
The LICENSE body: the Apache License, Version 2.0 text, with its appendix template left unfilled.
- https://raw.githubusercontent.com/THU-BPM/MarkLLM/main/README.md recorded 2026-11-12
README: "MarkLLM is an open-source toolkit developed to facilitate the research and application of watermarking technologies within large language models"; install from PyPI with "pip install markllm"; no pricing, paid tier or license key.
- https://ungh.cc/repos/THU-BPM/MarkLLM/files/main recorded 2026-11-12
The 195-file tree of THU-BPM/MarkLLM at `main`: config, dataset, evaluation, visualize and watermark folders, one root LICENSE, and no ee/, enterprise/, commercial/ or proprietary/ path. Path absence only, paired with the README cited beside it.
Adoption
1 medium confidencePyPI downloads of the markllm package, the toolkit's published install.
- https://pypi.org/pypi/markllm/json recorded 2026-11-12
PyPI markllm 0.1.5, uploaded 2024-10-21, summary "MarkLLM: An Open-Source Toolkit for LLM Watermarking", home page https://github.com/THU-BPM/MarkLLM.
- https://pypistats.org/api/packages/markllm/recent recorded 2026-11-12
last_month = 103 downloads for the package markllm.
Capability
2 medium confidenceA MarkLLM watermark is checked by running the toolkit's detector with the key the generator used, so an outsider who lacks the key takes the operator's word, as with SynthID-Text. UPV, one method among many, lets anyone holding its detector network check a mark; the toolkit's usual output is still a detection score that only the key holder can reproduce.
- https://arxiv.org/abs/2405.10051 recorded 2026-11-12
Abstract: "we introduce MarkLLM, an open-source toolkit for LLM watermarking" with "a comprehensive suite of 12 tools spanning three perspectives, along with two types of automated evaluation pipelines".
- https://raw.githubusercontent.com/THU-BPM/MarkLLM/main/config/KGW.json recorded 2026-11-12
The KGW configuration: "hash_key": 15485863, gamma, delta and "z_threshold": 4.0.
- https://raw.githubusercontent.com/THU-BPM/MarkLLM/main/config/UPV.json recorded 2026-11-12
The UPV configuration: "detect_mode": "network" with a separate "detector_model_name" beside the generator model.
- https://raw.githubusercontent.com/THU-BPM/MarkLLM/main/README.md recorded 2026-11-12
The supported-algorithm table (KGW, Unigram, SWEET, UPV, EWD, SIR, X-SIR, DiPmark, Unbiased, TS, SynthID-Text, PF, MorphMark, SemStamp and others), the visualization tools, and "12 evaluation tools that cover detectability, robustness, and impact on text quality".
Verified 2026-11-12