# coding=utf-8 # Copyright 2023 Authors of "A Watermark for Large Language Models" # available at https://arxiv.org/abs/2301.10226 # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import argparse from argparse import Namespace from pprint import pprint from functools import partial import numpy # for gradio hot reload import gradio as gr import torch from transformers import (AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForCausalLM, LogitsProcessorList) from watermark_processor import WatermarkLogitsProcessor, WatermarkDetector def str2bool(v): """Util function for user friendly boolean flag args""" if isinstance(v, bool): return v if v.lower() in ('yes', 'true', 't', 'y', '1'): return True elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False else: raise argparse.ArgumentTypeError('Boolean value expected.') def parse_args(): """Command line argument specification""" parser = argparse.ArgumentParser(description="A minimum working example of applying the watermark to any LLM that supports the huggingface 🤗 `generate` API") parser.add_argument( "--run_gradio", type=str2bool, default=True, help="Whether to launch as a gradio demo. Set to False if not installed and want to just run the stdout version.", ) parser.add_argument( "--demo_public", type=str2bool, default=False, help="Whether to expose the gradio demo to the internet.", ) parser.add_argument( "--model_name_or_path", type=str, default="facebook/opt-6.7b", help="Main model, path to pretrained model or model identifier from huggingface.co/models.", ) parser.add_argument( "--prompt_max_length", type=int, default=None, help="Truncation length for prompt, overrides model config's max length field.", ) parser.add_argument( "--max_new_tokens", type=int, default=200, help="Maximmum number of new tokens to generate.", ) parser.add_argument( "--generation_seed", type=int, default=123, help="Seed for setting the torch global rng prior to generation.", ) parser.add_argument( "--use_sampling", type=str2bool, default=True, help="Whether to generate using multinomial sampling.", ) parser.add_argument( "--sampling_temp", type=float, default=0.7, help="Sampling temperature to use when generating using multinomial sampling.", ) parser.add_argument( "--n_beams", type=int, default=1, help="Number of beams to use for beam search. 1 is normal greedy decoding", ) parser.add_argument( "--use_gpu", type=str2bool, default=True, help="Whether to run inference and watermark hashing/seeding/permutation on gpu.", ) parser.add_argument( "--seeding_scheme", type=str, default="simple_1", help="Seeding scheme to use to generate the greenlists at each generation and verification step.", ) parser.add_argument( "--gamma", type=float, default=0.25, help="The fraction of the vocabulary to partition into the greenlist at each generation and verification step.", ) parser.add_argument( "--delta", type=float, default=2.0, help="The amount/bias to add to each of the greenlist token logits before each token sampling step.", ) parser.add_argument( "--normalizers", type=str, default="", help="Single or comma separated list of the preprocessors/normalizer names to use when performing watermark detection.", ) parser.add_argument( "--ignore_repeated_bigrams", type=str2bool, default=False, help="Whether to use the detection method that only counts each unqiue bigram once as either a green or red hit.", ) parser.add_argument( "--detection_z_threshold", type=float, default=4.0, help="The test statistic threshold for the detection hypothesis test.", ) parser.add_argument( "--select_green_tokens", type=str2bool, default=True, help="How to treat the permuation when selecting the greenlist tokens at each step. Legacy is (False) to pick the complement/reds first.", ) parser.add_argument( "--skip_model_load", type=str2bool, default=False, help="Skip the model loading to debug the interface.", ) parser.add_argument( "--seed_separately", type=str2bool, default=True, help="Whether to call the torch seed function before both the unwatermarked and watermarked generate calls.", ) args = parser.parse_args() return args def load_model(args): """Load and return the model and tokenizer""" args.is_seq2seq_model = any([(model_type in args.model_name_or_path) for model_type in ["t5","T0"]]) args.is_decoder_only_model = any([(model_type in args.model_name_or_path) for model_type in ["gpt","opt","bloom"]]) if args.is_seq2seq_model: model = AutoModelForSeq2SeqLM.from_pretrained(args.model_name_or_path) elif args.is_decoder_only_model: model = AutoModelForCausalLM.from_pretrained(args.model_name_or_path) else: raise ValueError(f"Unknown model type: {args.model_name_or_path}") if args.use_gpu: device = "cuda" if torch.cuda.is_available() else "cpu" model = model.to(device) else: device = "cpu" model.eval() tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path) return model, tokenizer, device def generate(prompt, args, model=None, device=None, tokenizer=None): """Instatiate the WatermarkLogitsProcessor according to the watermark parameters and generate watermarked text by passing it to the generate method of the model as a logits processor. """ print(f"Generating with {args}") watermark_processor = WatermarkLogitsProcessor(vocab=list(tokenizer.get_vocab().values()), gamma=args.gamma, delta=args.delta, seeding_scheme=args.seeding_scheme, select_green_tokens=args.select_green_tokens) gen_kwargs = dict(max_new_tokens=args.max_new_tokens) if args.use_sampling: gen_kwargs.update(dict( do_sample=True, top_k=0, temperature=args.sampling_temp )) else: gen_kwargs.update(dict( num_beams=args.n_beams )) generate_without_watermark = partial( model.generate, **gen_kwargs ) generate_with_watermark = partial( model.generate, logits_processor=LogitsProcessorList([watermark_processor]), **gen_kwargs ) if args.prompt_max_length: pass elif hasattr(model.config,"max_position_embedding"): args.prompt_max_length = model.config.max_position_embeddings-args.max_new_tokens else: args.prompt_max_length = 2048-args.max_new_tokens tokd_input = tokenizer(prompt, return_tensors="pt", add_special_tokens=True, truncation=True, max_length=args.prompt_max_length).to(device) truncation_warning = True if tokd_input["input_ids"].shape[-1] == args.prompt_max_length else False redecoded_input = tokenizer.batch_decode(tokd_input["input_ids"], skip_special_tokens=True)[0] torch.manual_seed(args.generation_seed) output_without_watermark = generate_without_watermark(**tokd_input) # optional to seed before second generation, but will not be the same again generally, unless delta==0.0, no-op watermark if args.seed_separately: torch.manual_seed(args.generation_seed) output_with_watermark = generate_with_watermark(**tokd_input) if args.is_decoder_only_model: # need to isolate the newly generated tokens output_without_watermark = output_without_watermark[:,tokd_input["input_ids"].shape[-1]:] output_with_watermark = output_with_watermark[:,tokd_input["input_ids"].shape[-1]:] decoded_output_without_watermark = tokenizer.batch_decode(output_without_watermark, skip_special_tokens=True)[0] decoded_output_with_watermark = tokenizer.batch_decode(output_with_watermark, skip_special_tokens=True)[0] return (redecoded_input, int(truncation_warning), decoded_output_without_watermark, decoded_output_with_watermark, args) # decoded_output_with_watermark) def format_names(s): """Format names for the gradio demo interface""" s=s.replace("num_tokens_scored","Tokens Counted (T)") s=s.replace("num_green_tokens","# Tokens in Greenlist") s=s.replace("green_fraction","Fraction of T in Greenlist") s=s.replace("z_score","z-score") s=s.replace("p_value","p value") return s def list_format_scores(score_dict, detection_threshold): """Format the detection metrics into a gradio dataframe input format""" lst_2d = [] lst_2d.append(["z-score threshold", f"{detection_threshold}"]) for k,v in score_dict.items(): if k=='green_fraction': lst_2d.append([format_names(k), f"{v:.1%}"]) elif k=='confidence': lst_2d.append([format_names(k), f"{v:.3%}"]) elif isinstance(v, float): lst_2d.append([format_names(k), f"{v:.3g}"]) elif isinstance(v, bool): lst_2d.append([format_names(k), ("Watermarked" if v else "Human/Unwatermarked")]) else: lst_2d.append([format_names(k), f"{v}"]) return lst_2d def detect(input_text, args, device=None, tokenizer=None): """Instantiate the WatermarkDetection object and call detect on the input text returning the scores and outcome of the test""" watermark_detector = WatermarkDetector(vocab=list(tokenizer.get_vocab().values()), gamma=args.gamma, seeding_scheme=args.seeding_scheme, device=device, tokenizer=tokenizer, z_threshold=args.detection_z_threshold, normalizers=args.normalizers, ignore_repeated_bigrams=args.ignore_repeated_bigrams, select_green_tokens=args.select_green_tokens) if len(input_text)-1 > watermark_detector.min_prefix_len: score_dict = watermark_detector.detect(input_text) # output = str_format_scores(score_dict, watermark_detector.z_threshold) output = list_format_scores(score_dict, watermark_detector.z_threshold) else: # output = (f"Error: string not long enough to compute watermark presence.") output = [["Error","string too short to compute metrics"]] output += [["",""] for _ in range(6)] return output, args def run_gradio(args, model=None, device=None, tokenizer=None): """Define and launch the gradio demo interface""" generate_partial = partial(generate, model=model, device=device, tokenizer=tokenizer) detect_partial = partial(detect, device=device, tokenizer=tokenizer) with gr.Blocks() as demo: # Top section, greeting and instructions gr.Markdown("## 💧 [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226) 🔍") gr.Markdown("[jwkirchenbauer/lm-watermarking![](https://badgen.net/badge/icon/GitHub?icon=github&label)](https://github.com/jwkirchenbauer/lm-watermarking)") gr.Markdown(f"Language model: {args.model_name_or_path}") with gr.Accordion("Understanding the output metrics",open=False): gr.Markdown( """ - `z-score threshold` : The cuttoff for the hypothesis test - `Tokens Counted (T)` : The number of tokens in the output that were counted by the detection algorithm. The first token is ommitted in the simple, single token seeding scheme since there is no way to generate a greenlist for it as it has no prefix token(s). Under the "Ignore Bigram Repeats" detection algorithm, described in the bottom panel, this can be much less than the total number of tokens generated if there is a lot of repetition. - `# Tokens in Greenlist` : The number of tokens that were observed to fall in their respective greenlist - `Fraction of T in Greenlist` : The `# Tokens in Greenlist` / `T`. This is expected to be approximately `gamma` for human/unwatermarked text. - `z-score` : The test statistic for the detection hypothesis test. If larger than the `z-score threshold` we "reject the null hypothesis" that the text is human/unwatermarked, and conclude it is watermarked - `p value` : The likelihood of observing the computed `z-score` under the null hypothesis. This is the likelihood of observing the `Fraction of T in Greenlist` given that the text was generated without knowledge of the watermark procedure/greenlists. If this is extremely _small_ we are confident that this many green tokens was not chosen by random chance. - `prediction` : The outcome of the hypothesis test - whether the observed `z-score` was higher than the `z-score threshold` - `confidence` : If we reject the null hypothesis, and the `prediction` is "Watermarked", then we report 1-`p value` to represent the confidence of the detection based on the unlikeliness of this `z-score` observation. """ ) with gr.Accordion("A note on model capability",open=False): gr.Markdown( """ The models that can be used in this demo are limited to those that are both open source and that fit on a single commodity GPU. In particular, there aren't many models above a few billion parameters and almost none trained using both Instruction-finetuning an/or RLHF. Therefore, in both it's un-watermarked (normal) and watermarked states, the model is not generally able to respond well to the kinds of prompts that a 100B+ Instruction and RLHF tuned model such as ChatGPT, Claude, or Bard is. We suggest you try prompts that give the model a few sentences and then allow it to 'continue' the prompt, as these weaker models are more capable in this simpler language modeling setting. Some examples include the opening paragraph of a wikipedia article, or the first few sentences of a story. Longer prompts and stopping mid sentence often helps encourage more fluent, longer genrations. """ ) # Construct state for parameters, define updates and toggles session_args = gr.State(value=args) with gr.Tab("Generate and Detect"): with gr.Row(): prompt = gr.Textbox(label=f"Prompt", interactive=True,lines=10,max_lines=10) with gr.Row(): generate_btn = gr.Button("Generate") with gr.Row(): with gr.Column(scale=2): output_without_watermark = gr.Textbox(label="Output Without Watermark", interactive=False,lines=14,max_lines=14) with gr.Column(scale=1): # without_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14) without_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2) with gr.Row(): with gr.Column(scale=2): output_with_watermark = gr.Textbox(label="Output With Watermark", interactive=False,lines=14,max_lines=14) with gr.Column(scale=1): # with_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14) with_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"],interactive=False,row_count=7,col_count=2) redecoded_input = gr.Textbox(visible=False) truncation_warning = gr.Number(visible=False) def truncate_prompt(redecoded_input, truncation_warning, orig_prompt, args): if truncation_warning: return redecoded_input + f"\n\n[Prompt was truncated before generation due to length...]", args else: return orig_prompt, args with gr.Tab("Detector Only"): with gr.Row(): with gr.Column(scale=2): detection_input = gr.Textbox(label="Text to Analyze", interactive=True,lines=14,max_lines=14) with gr.Column(scale=1): # detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14) detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2) with gr.Row(): detect_btn = gr.Button("Detect") # Parameter selection group with gr.Accordion("Advanced Settings",open=False): with gr.Row(): with gr.Column(scale=1): gr.Markdown(f"#### Generation Parameters") with gr.Row(): decoding = gr.Radio(label="Decoding Method",choices=["multinomial", "greedy"], value=("multinomial" if args.use_sampling else "greedy")) with gr.Row(): sampling_temp = gr.Slider(label="Sampling Temperature", minimum=0.1, maximum=1.0, step=0.1, value=args.sampling_temp, visible=True) with gr.Row(): generation_seed = gr.Number(label="Generation Seed",value=args.generation_seed, interactive=True) with gr.Row(): n_beams = gr.Dropdown(label="Number of Beams",choices=list(range(1,11,1)), value=args.n_beams, visible=(not args.use_sampling)) with gr.Row(): max_new_tokens = gr.Slider(label="Max Generated Tokens", minimum=10, maximum=1000, step=10, value=args.max_new_tokens) with gr.Column(scale=1): gr.Markdown(f"#### Watermark Parameters") with gr.Row(): gamma = gr.Slider(label="gamma",minimum=0.1, maximum=0.9, step=0.05, value=args.gamma) with gr.Row(): delta = gr.Slider(label="delta",minimum=0.0, maximum=10.0, step=0.1, value=args.delta) gr.Markdown(f"#### Detector Parameters") with gr.Row(): detection_z_threshold = gr.Slider(label="z-score threshold",minimum=0.0, maximum=10.0, step=0.1, value=args.detection_z_threshold) with gr.Row(): ignore_repeated_bigrams = gr.Checkbox(label="Ignore Bigram Repeats") with gr.Row(): normalizers = gr.CheckboxGroup(label="Normalizations", choices=["unicode", "homoglyphs", "truecase"], value=args.normalizers) # with gr.Accordion("Actual submitted parameters:",open=False): with gr.Row(): gr.Markdown(f"_Note: sliders don't always update perfectly. Clicking on the bar or using the number window to the right can help. Window below shows the current settings._") with gr.Row(): current_parameters = gr.Textbox(label="Current Parameters", value=args) with gr.Accordion("Legacy Settings",open=False): with gr.Row(): with gr.Column(scale=1): seed_separately = gr.Checkbox(label="Seed both generations separately", value=args.seed_separately) with gr.Column(scale=1): select_green_tokens = gr.Checkbox(label="Select 'greenlist' from partition", value=args.select_green_tokens) with gr.Accordion("Understanding the settings",open=False): gr.Markdown( """ #### Generation Parameters: - Decoding Method : We can generate tokens from the model using either multinomial sampling or we can generate using greedy decoding. - Sampling Temperature : If using multinomial sampling we can set the temperature of the sampling distribution. 0.0 is equivalent to greedy decoding, and 1.0 is the maximum amount of variability/entropy in the next token distribution. 0.7 strikes a nice balance between faithfulness to the model's estimate of top candidates while adding variety. Does not apply for greedy decoding. - Generation Seed : The integer to pass to the torch random number generator before running generation. Makes the multinomial sampling strategy outputs reproducible. Does not apply for greedy decoding. - Number of Beams : When using greedy decoding, we can also set the number of beams to > 1 to enable beam search. This is not implemented/excluded from paper for multinomial sampling but may be added in future. - Max Generated Tokens : The `max_new_tokens` parameter passed to the generation method to stop the output at a certain number of new tokens. Note that the model is free to generate fewer tokens depending on the prompt. Implicitly this sets the maximum number of prompt tokens possible as the model's maximum input length minus `max_new_tokens`, and inputs will be truncated accordingly. #### Watermark Parameters: - gamma : The fraction of the vocabulary to be partitioned into the greenlist at each generation step. Smaller gamma values create a stronger watermark by enabling the watermarked model to achieve a greater differentiation from human/unwatermarked text because it is preferentially sampling from a smaller green set making those tokens less likely to occur by chance. - delta : The amount of positive bias to add to the logits of every token in the greenlist at each generation step before sampling/choosing the next token. Higher delta values mean that the greenlist tokens are more heavily preferred by the watermarked model and as the bias becomes very large the watermark transitions from "soft" to "hard". For a hard watermark, nearly all tokens are green, but this can have a detrimental effect on generation quality, especially when there is not a lot of flexibility in the distribution. #### Detector Parameters: - z-score threshold : the z-score cuttoff for the hypothesis test. Higher thresholds (such as 4.0) make _false positives_ (predicting that human/unwatermarked text is watermarked) very unlikely as a genuine human text with a significant number of tokens will almost never achieve that high of a z-score. Lower thresholds will capture more _true positives_ as some watermarked texts will contain less green tokens and achive a lower z-score, but still pass the lower bar and be flagged as "watermarked". However, a lowere threshold will increase the chance that human text that contains a slightly higher than average number of green tokens is erroneously flagged. 4.0-5.0 offers extremely low false positive rates while still accurately catching most watermarked text. - Ignore Bigram Repeats : This alternate detection algorithm only considers the unique bigrams in the text during detection, computing the greenlists based on the first in each pair and checking whether the second falls within the list. This means that `T` is now the unique number of bigrams in the text, which becomes less than the total number of tokens generated if the text contains a lot of repetition. See the paper for a more detailed discussion. - Normalizations : we implement a few basic normaliations to defend against various adversarial perturbations of the text analyzed during detection. Currently we support converting all chracters to unicode, replacing homoglyphs with a canonical form, and standardizing the capitalization. See the paper for a detailed discussion of input normalization. """ ) gr.HTML("""
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
Follow the github link at the top and host the demo on your own GPU hardware to test out larger models.