How LLM Watermarking Works

ModeliThe language model that generates the text, running fully in your browser. Smaller models download faster and need less memory; larger ones write better prose. Weights are fetched once from the Hugging Face CDN and cached.

Idle. WebGPU

PromptiWhat you want the model to write about. In Instruction mode it is treated as a chat request; in Continuation mode the model simply continues your text.

Watermarking modeiHow the mark is embedded. None: no watermark. Hard: red tokens are banned outright. Soft: green tokens get a small score boost. Tournament: candidates compete on a keyed g-value. See "How it works" for the full picture.

Watermarking parameters

γ fractioniShare of the vocabulary placed on the green list at each step. 0.5 means half the tokens are green. Lower values make the mark sparser but each green token carries more signal.

δ biasiHow much to boost the score of green tokens (soft mode). Higher δ makes the watermark easier to detect but can hurt fluency, since the model is pushed harder away from its natural choice.

h contextiHow many preceding tokens seed the pseudo-random green/red split. The same h tokens plus the same key always reproduce the same split, which is how the detector replays it.

keyiThe secret that seeds the watermark. Detection only succeeds with the matching key: change it here versus the detector to watch the evidence collapse to chance. Force onto the red list (optional, comma-separated)iWords you want to force onto the red list regardless of the seed. Handy for demos: forbid a word in hard mode and watch the model route around it.

Generation parameters

max tokensiUpper bound on how many tokens to generate. Longer outputs give the detector more evidence, so a real watermark reads as more confident.

temperatureiRandomness of sampling. Higher is more creative and varied; lower is more focused and repetitive. Around 0.7 is a common balance.

top-piNucleus sampling: only consider the smallest set of tokens whose probabilities sum to p. 0.95 keeps the plausible options and drops the long tail.

top-kiOnly sample from the k most likely tokens at each step. Smaller k is safer and more predictable; larger k allows more variety.

rep. penaltyiDiscourages repeating tokens already generated. Above 1.0 reduces loops and repeated phrases; too high can make text disjointed.

New here? Read how the three modes work.

Output

Pick a model and press Generate.

Detection

Detector keyiThe key the detector uses to replay the green/red split. It must match the generation key to find the mark. Try a wrong key to see detection fall back to chance.