WatermarkedCausalLanguageModel<T>

Models & Types · AiDotNet.Safety.Watermarking

A language model whose next-token logits carry a generation-time watermark: the inner model’s logits plus the watermarker’s green-list bias for the current context.

Implements: ICausalLanguageModel<T>

Known derived types: WatermarkedIncrementalCausalLanguageModel<T>

For Beginners

this wraps your model so everything it writes carries your watermark, without changing the model itself.

How It Works

Created by WatermarkModel. The inner model’s logits are copied before the bias is added, so a model that reuses its output buffer is not modified. Use the wrapped model anywhere an ICausalLanguageModel is accepted, for example in LocalEngineChatClient; every token it samples is then drawn from the watermarked distribution.

Constructors

Constructor Summary
WatermarkedCausalLanguageModel(ICausalLanguageModel<T> inner, IGenerationWatermarker<T> watermarker) Wraps inner with watermarker’s logit bias.

Properties

Property Summary
IGenerationWatermarker<T> Watermarker { get; } The watermarker whose bias is applied.
int VocabularySize { get; }

Methods

Method Summary
Vector<T> NextTokenLogits(IReadOnlyList<int> tokenIds)