LLM Watermarking from Scratch (Gist)

protagonist_hn1 pts1 comments

Minimal LLM Watermarking from scratch · GitHub

/" data-turbo-transient="true" />

Skip to content

-->

Search Gists

Search Gists

Sign in

Sign up

You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

{{ message }}

Instantly share code, notes, and snippets.

jSwords91/minimark

Created<br>August 23, 2026 15:47

Show Gist options

Download ZIP

Star

(0)

You must be signed in to star a gist

Fork

(0)

You must be signed in to fork a gist

Embed

Select an option

Embed<br>Embed this gist in your website.

Share<br>Copy sharable link for this gist.

Clone via HTTPS<br>Clone using the web URL.

No results found

Learn more about clone URLs

Clone this repository at &lt;script src=&quot;https://gist.github.com/jSwords91/2732ff6017213526da73e8dc2bd54770.js&quot;&gt;&lt;/script&gt;

" readonly="readonly" data-autoselect="true" data-target="primer-text-field.inputElement " aria-describedby="validation-6fe28768-1110-4a32-a912-3e5e39cdc2d9" class="form-control FormControl-monospace FormControl-input FormControl-small rounded-left-0 rounded-right-0 border-right-0" type="text" name="gist-share-url-sized-down" />

Save jSwords91/2732ff6017213526da73e8dc2bd54770 to your computer and use it in GitHub Desktop.

Embed

Select an option

Embed<br>Embed this gist in your website.

Share<br>Copy sharable link for this gist.

Clone via HTTPS<br>Clone using the web URL.

No results found

Learn more about clone URLs

Clone this repository at &lt;script src=&quot;https://gist.github.com/jSwords91/2732ff6017213526da73e8dc2bd54770.js&quot;&gt;&lt;/script&gt;

" readonly="readonly" data-autoselect="true" data-target="primer-text-field.inputElement " aria-describedby="validation-608a3428-d4dd-4850-a218-8c73da20f715" class="form-control FormControl-monospace FormControl-input FormControl-small rounded-left-0 rounded-right-0 border-right-0" type="text" name="gist-share-url-original" />

Save jSwords91/2732ff6017213526da73e8dc2bd54770 to your computer and use it in GitHub Desktop.

Download ZIP

Minimal LLM Watermarking from scratch

Raw

minimark

This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.<br>Learn more about bidirectional Unicode characters

Show hidden characters

#!/usr/bin/env -S uv run --script

# /// script

# requires-python = ">=3.11"

# dependencies = [

# "torch",

# "transformers>=5.15",

# ]

# ///

"""LLM Watermarks.

Anthropic will roll out their "is it AI" API soon.

This is ~roughly how it works.

SynthID-Text from scratch.

A tiny implementation of Tournament Sampling + watermark detection.

uv run synthid.py

"""

from __future__ import annotations

import hashlib

import hmac

import math

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Qwen/Qwen2.5-0.5B-Instruct"

KEY = b"player-piano"

H = 4 # previous tokens used as watermark context

M = 30 # tournament layers

TOP_K = 100

TEMP = 0.7

MAX_NEW = 120

def bits(ctx: tuple[int, ...], tok: int, key: bytes = KEY) -> list[int]:

"""Return M deterministic keyed random bits for (context, token)."""

msg = b"".join(int(x).to_bytes(4, "little") for x in (*ctx, tok))

digest = hmac.new(key, msg, hashlib.sha256).digest()

n = int.from_bytes(digest, "little")

return [(n >> i) & 1 for i in range(M)]

def watermark(p: torch.Tensor, toks: torch.Tensor, ctx: tuple[int, ...]) -> torch.Tensor:

"""Apply M layers of N=2 Tournament Sampling."""

ids = toks.detach().cpu().tolist()

g = torch.tensor([bits(ctx, t) for t in ids], dtype=torch.float32, device=p.device)

p = p.float()

for i in range(M):

q = (p * g[:, i]).sum().clamp(0, 1)

# The whole trick:

# p'(x) = p(x) [1 + g(x) - q]

p = p * (1 + g[:, i] - q)

p = p.clamp_min(0)

return p / p.sum()

@torch.inference_mode()

def generate(model, tok, prompt: str, *, wm: bool, seed: int = 42) -> tuple[str, list[int]]:

"""Generate text with or without the watermark."""

device = next(model.parameters()).device

torch.manual_seed(seed)

chat = tok.apply_chat_template(

[{"role": "user", "content": prompt}],

tokenize=False,

add_generation_prompt=True,

prompt_ids = tok(chat, return_tensors="pt", add_special_tokens=False).input_ids.to(device)

out: list[int] = []

seen: set[tuple[int, ...]] = set()

eos = model.generation_config.eos_token_id

if eos is None:

eos = tok.eos_token_id

eos_ids = set(eos if isinstance(eos, (list, tuple)) else [eos])

for _ in range(MAX_NEW):

ids = prompt_ids

if out:

ids = torch.cat(

[prompt_ids, torch.tensor([out], dtype=torch.long, device=device)],

dim=1,

logits = model(ids, use_cache=False).logits[0, -1].float() / TEMP

top_logits, top_ids = torch.topk(logits, min(TOP_K, logits.numel()))

p = torch.softmax(top_logits, dim=-1)

#...

torch gist clone text device from

Related Articles