# coding=utf-8 # Copyright 2023 Anonymous Authors of "A Watermark for Large Language Models" # # 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 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 transformers import GPT2TokenizerFast OPT_TOKENIZER = GPT2TokenizerFast from watermark_processor import WatermarkLogitsProcessor, WatermarkDetector API_MODEL_MAP = { "bigscience/bloom" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0}, "bigscience/bloomz" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0}, "google/flan-ul2" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0}, "google/flan-t5-xxl" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0}, "EleutherAI/gpt-neox-20b" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0}, } 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.", ) parser.add_argument( "--load_fp16", type=str2bool, default=False, help="Whether to run model in float16 precsion.", ) 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: if args.load_fp16: model = AutoModelForCausalLM.from_pretrained(args.model_name_or_path,torch_dtype=torch.float16, device_map='auto') else: 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" if args.load_fp16: pass else: model = model.to(device) else: device = "cpu" model.eval() tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path) return model, tokenizer, device from text_generation import InferenceAPIClient from requests.exceptions import ReadTimeout def generate_with_api(prompt, args): hf_api_key = os.environ.get("HF_API_KEY") if hf_api_key is None: raise ValueError("HF_API_KEY environment variable not set, cannot use HF API to generate text.") client = InferenceAPIClient(args.model_name_or_path, token=hf_api_key, timeout=60) assert args.n_beams == 1, "HF API models do not support beam search." generation_params = { "max_new_tokens": args.max_new_tokens, "do_sample": args.use_sampling, } if args.use_sampling: generation_params["temperature"] = args.sampling_temp generation_params["seed"] = args.generation_seed timeout_msg = "[Model API timeout error. Try reducing the max_new_tokens parameter or the prompt length.]" try: generation_params["watermark"] = False without_watermark_iterator = client.generate_stream(prompt, **generation_params) except ReadTimeout as e: print(e) without_watermark_iterator = (char for char in timeout_msg) try: generation_params["watermark"] = True with_watermark_iterator = client.generate_stream(prompt, **generation_params) except ReadTimeout as e: print(e) with_watermark_iterator = (char for char in timeout_msg) all_without_words, all_with_words = "", "" for without_word, with_word in zip(without_watermark_iterator, with_watermark_iterator): all_without_words += without_word.token.text all_with_words += with_word.token.text yield all_without_words, all_with_words def check_prompt(prompt, args, tokenizer, model=None, device=None): # This applies to both the local and API model scenarios if args.model_name_or_path in API_MODEL_MAP: args.prompt_max_length = API_MODEL_MAP[args.model_name_or_path]["max_length"] 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] return (redecoded_input, int(truncation_warning), args) def generate(prompt, args, tokenizer, model=None, device=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}") print(f"Prompt: {prompt}") if args.model_name_or_path in API_MODEL_MAP: api_outputs = generate_with_api(prompt, args) yield from api_outputs else: tokd_input = tokenizer(prompt, return_tensors="pt", add_special_tokens=True, truncation=True, max_length=args.prompt_max_length).to(device) 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 ) 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] # mocking the API outputs in a whitespace split generator style all_without_words, all_with_words = "", "" for without_word, with_word in zip(decoded_output_without_watermark.split(), decoded_output_with_watermark.split()): all_without_words += without_word + " " all_with_words += with_word + " " yield all_without_words, all_with_words 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") s=s.replace("prediction","Prediction") s=s.replace("confidence","Confidence") return s def list_format_scores(score_dict, detection_threshold): """Format the detection metrics into a gradio dataframe input format""" lst_2d = [] 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}"]) if "confidence" in score_dict: lst_2d.insert(-2,["z-score Threshold", f"{detection_threshold}"]) else: lst_2d.insert(-1,["z-score Threshold", f"{detection_threshold}"]) return lst_2d def detect(input_text, args, tokenizer, device=None, return_green_token_mask=True): """Instantiate the WatermarkDetection object and call detect on the input text returning the scores and outcome of the test""" print(f"Detecting with {args}") print(f"Detection Tokenizer: {type(tokenizer)}") 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) # for now, just don't display the green token mask # if we're using normalizers or ignore_repeated_bigrams if args.normalizers != [] or args.ignore_repeated_bigrams: return_green_token_mask = False error = False green_token_mask = None if input_text == "": error = True else: try: score_dict = watermark_detector.detect(input_text, return_green_token_mask=return_green_token_mask) green_token_mask = score_dict.pop("green_token_mask", None) output = list_format_scores(score_dict, watermark_detector.z_threshold) except ValueError as e: print(e) error = True if error: output = [["Error","string too short to compute metrics"]] output += [["",""] for _ in range(6)] html_output = "[No highlight markup generated]" if green_token_mask is not None: # hack bc we need a fast tokenizer with charspan support if "opt" in args.model_name_or_path: tokenizer = OPT_TOKENIZER.from_pretrained(args.model_name_or_path) tokens = tokenizer(input_text) if tokens["input_ids"][0] == tokenizer.bos_token_id: tokens["input_ids"] = tokens["input_ids"][1:] # ignore attention mask skip = watermark_detector.min_prefix_len charspans = [tokens.token_to_chars(i) for i in range(skip,len(tokens["input_ids"]))] charspans = [cs for cs in charspans if cs is not None] # remove the special token spans if len(charspans) != len(green_token_mask): breakpoint() assert len(charspans) == len(green_token_mask) tags = [(f'{input_text[cs.start:cs.end]}' if m else f'{input_text[cs.start:cs.end]}') for cs, m in zip(charspans, green_token_mask)] html_output = f'

{" ".join(tags)}

' return output, args, tokenizer, html_output def run_gradio(args, model=None, device=None, tokenizer=None): """Define and launch the gradio demo interface""" check_prompt_partial = partial(check_prompt, model=model, device=device) generate_partial = partial(generate, model=model, device=device) detect_partial = partial(detect, device=device) css = """ .green { color: black!important;line-height:1.9em; padding: 0.2em 0.2em; background: #ccffcc; border-radius:0.5rem;} .red { color: black!important;line-height:1.9em; padding: 0.2em 0.2em; background: #ffad99; border-radius:0.5rem;} """ with gr.Blocks(css=css) as demo: # Top section, greeting and instructions with gr.Row(): with gr.Column(scale=9): gr.Markdown( """ ## 💧 A Watermark for Large Language Models 🔍 Demo made possible by the HuggingFace 🤗 [text-generation-inference](https://github.com/huggingface/text-generation-inference) serving framework. """ ) with gr.Column(scale=1): # if model_name_or_path at startup not one of the API models then add to dropdown all_models = sorted(list(set(list(API_MODEL_MAP.keys())+[args.model_name_or_path]))) model_selector = gr.Dropdown( all_models, value=args.model_name_or_path, label="Language Model", ) # Construct state for parameters, define updates and toggles default_prompt = args.__dict__.pop("default_prompt") session_args = gr.State(value=args) # note that state obj automatically calls value if it's a callable, want to avoid calling tokenizer at startup session_tokenizer = gr.State(value=lambda : tokenizer) with gr.Tab("Welcome"): with gr.Row(): with gr.Column(scale=2): gr.Markdown( """ Potential harms of large language models can be mitigated by *watermarking* a model's output. *Watermarks* are embedded signals in the generated text that are invisible to humans but algorithmically detectable, that allow *anyone* to later check whether a given span of text was likely to have been generated by a model that uses the watermark. This space showcases a watermarking approach that can be applied to _any_ generative language model. For demonstration purposes, the space demos a selection of multi-billion parameter models (see the following note for caveats). """ ) with gr.Accordion("A note on the available models:",open=False): gr.Markdown( """ This demo uses open-source language models. Today, these models are less powerful than proprietary commercial tools like ChatGPT, Claude, Bard, or Bing/Sydney. Models like [BLOOM (175B)](https://huggingface.co/bigscience/bloom) are designed to "complete" your prompt, and are not fine-tuned to follow instructions. For best results, prompt that model with a few sentences that form the beginning of a paragraph, and then allow it to "continue" your paragraph. Some examples include the opening paragraph of a wikipedia article, or the first few sentences of a story. Longer prompts that end mid-sentence will result in more fluent generations. Some of the models available in this demo are fine-tuned to follow instructions but have different strengths and will showcase differenttypes of watermark behavior. [BLOOMZ](https://huggingface.co/bigscience/bloomz) is an instruction tuned variant of BLOOM capable of following instructions in dozens of languages zero-shot and can generate long and coherent paragraphs and stories given the right prompt. The FLAN models [FLAN-t5-xxl (11B)](https://huggingface.co/google/flan-t5-xxl) and [FLAN-UL2 (20B)](https://huggingface.co/google/flan-ul2) are fine-tuned on a variety of in-context few-shot learning NLP tasks, such as reasoning, and question answering. Generally, short, low entropy scenarios where the model has very few choices in terms of correct/suitable responses to the prompt will not exhibit as strong of a watermark presence, while longer watermarked outputs will produce higher detection statistics. """ ) gr.Markdown( """ **[Generate & Detect]**: The first tab shows that the watermark can be embedded with negligible impact on text quality. You can try any prompt and compare the quality of normal text (*Output Without Watermark*) to the watermarked text (*Output With Watermark*) below it. You can also "see" the watermark by looking at the **Highlighted** tab where the tokens are colored green or red depending on which list they are in. Metrics on the right show that the watermark can be reliably detected given a reasonably small number of tokens (25-50). Detection is very efficient and does not use the language model or its parameters. **[Detector Only]**: You can also copy-paste the watermarked text (or any other text) into the second tab. This can be used to see how many sentences you could remove and still detect the watermark. You can also verify here that the detection has, by design, a low false-positive rate; This means that human-generated text that you copy into this detector will not be marked as machine-generated. """ ) with gr.Column(scale=1): gr.Markdown( """ ![](https://drive.google.com/uc?export=view&id=1yVLPcjm-xvaCjQyc3FGLsWIU84v1QRoC) """ ) with gr.Tab("Generate & Detect"): with gr.Row(): prompt = gr.Textbox(label=f"Prompt", interactive=True,lines=10,max_lines=10, value=default_prompt) with gr.Row(): generate_btn = gr.Button("Generate") with gr.Row(): with gr.Column(scale=2): with gr.Tab("Output Without Watermark (Raw Text)"): output_without_watermark = gr.Textbox(interactive=False,lines=14,max_lines=14) with gr.Tab("Highlighted"): html_without_watermark = gr.HTML(elem_id="html-without-watermark") with gr.Column(scale=1): 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): with gr.Tab("Output With Watermark (Raw Text)"): output_with_watermark = gr.Textbox(interactive=False,lines=14,max_lines=14) with gr.Tab("Highlighted"): html_with_watermark = gr.HTML(elem_id="html-with-watermark") with gr.Column(scale=1): 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): with gr.Tab("Text to Analyze"): detection_input = gr.Textbox(interactive=True,lines=14,max_lines=14) with gr.Tab("Highlighted"): html_detection_input = gr.HTML(elem_id="html-detection-input") with gr.Column(scale=1): 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) and (not args.model_name_or_path in API_MODEL_MAP))) 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.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("What do the settings do?",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. """ ) with gr.Accordion("What do the output metrics mean?",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. """ ) # Register main generation tab click, outputing generations as well as a the encoded+redecoded+potentially truncated prompt and flag, then call detection generate_btn.click(fn=check_prompt_partial, inputs=[prompt,session_args,session_tokenizer], outputs=[redecoded_input, truncation_warning, session_args]).success( fn=generate_partial, inputs=[redecoded_input,session_args,session_tokenizer], outputs=[output_without_watermark, output_with_watermark]).success( fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]).success( fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) # Show truncated version of prompt if truncation occurred redecoded_input.change(fn=truncate_prompt, inputs=[redecoded_input,truncation_warning,prompt,session_args], outputs=[prompt,session_args]) # Register main detection tab click detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result, session_args,session_tokenizer,html_detection_input], api_name="detection") # State management logic # define update callbacks that change the state dict def update_model(session_state, value): session_state.model_name_or_path = value; return session_state def update_sampling_temp(session_state, value): session_state.sampling_temp = float(value); return session_state def update_generation_seed(session_state, value): session_state.generation_seed = int(value); return session_state def update_gamma(session_state, value): session_state.gamma = float(value); return session_state def update_delta(session_state, value): session_state.delta = float(value); return session_state def update_detection_z_threshold(session_state, value): session_state.detection_z_threshold = float(value); return session_state def update_decoding(session_state, value): if value == "multinomial": session_state.use_sampling = True elif value == "greedy": session_state.use_sampling = False return session_state def toggle_sampling_vis(value): if value == "multinomial": return gr.update(visible=True) elif value == "greedy": return gr.update(visible=False) def toggle_sampling_vis_inv(value): if value == "multinomial": return gr.update(visible=False) elif value == "greedy": return gr.update(visible=True) # if model name is in the list of api models, set the num beams parameter to 1 and hide n_beams def toggle_vis_for_api_model(value): if value in API_MODEL_MAP: return gr.update(visible=False) else: return gr.update(visible=True) def toggle_beams_for_api_model(value, orig_n_beams): if value in API_MODEL_MAP: return gr.update(value=1) else: return gr.update(value=orig_n_beams) # if model name is in the list of api models, set the interactive parameter to false def toggle_interactive_for_api_model(value): if value in API_MODEL_MAP: return gr.update(interactive=False) else: return gr.update(interactive=True) # if model name is in the list of api models, set gamma and delta based on API map def toggle_gamma_for_api_model(value, orig_gamma): if value in API_MODEL_MAP: return gr.update(value=API_MODEL_MAP[value]["gamma"]) else: return gr.update(value=orig_gamma) def toggle_delta_for_api_model(value, orig_delta): if value in API_MODEL_MAP: return gr.update(value=API_MODEL_MAP[value]["delta"]) else: return gr.update(value=orig_delta) def update_n_beams(session_state, value): session_state.n_beams = value; return session_state def update_max_new_tokens(session_state, value): session_state.max_new_tokens = int(value); return session_state def update_ignore_repeated_bigrams(session_state, value): session_state.ignore_repeated_bigrams = value; return session_state def update_normalizers(session_state, value): session_state.normalizers = value; return session_state def update_seed_separately(session_state, value): session_state.seed_separately = value; return session_state def update_select_green_tokens(session_state, value): session_state.select_green_tokens = value; return session_state def update_tokenizer(model_name_or_path): # if model_name_or_path == ALPACA_MODEL_NAME: # return ALPACA_MODEL_TOKENIZER.from_pretrained(ALPACA_TOKENIZER_PATH) # else: return AutoTokenizer.from_pretrained(model_name_or_path) def check_model(value): return value if (value!="" and value is not None) else args.model_name_or_path # enforce constraint that model cannot be null or empty # then attach model callbacks in particular model_selector.change(check_model, inputs=[model_selector], outputs=[model_selector]).then( toggle_vis_for_api_model,inputs=[model_selector], outputs=[n_beams] ).then( toggle_beams_for_api_model,inputs=[model_selector,n_beams], outputs=[n_beams] ).then( toggle_interactive_for_api_model,inputs=[model_selector], outputs=[gamma] ).then( toggle_interactive_for_api_model,inputs=[model_selector], outputs=[delta] ).then( toggle_gamma_for_api_model,inputs=[model_selector,gamma], outputs=[gamma] ).then( toggle_delta_for_api_model,inputs=[model_selector,delta], outputs=[delta] ).then( update_tokenizer,inputs=[model_selector], outputs=[session_tokenizer] ).then( update_model,inputs=[session_args, model_selector], outputs=[session_args] ).then( lambda value: str(value), inputs=[session_args], outputs=[current_parameters] ) # registering callbacks for toggling the visibilty of certain parameters based on the values of others decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[sampling_temp]) decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[generation_seed]) decoding.change(toggle_sampling_vis_inv,inputs=[decoding], outputs=[n_beams]) decoding.change(toggle_vis_for_api_model,inputs=[model_selector], outputs=[n_beams]) # registering all state update callbacks decoding.change(update_decoding,inputs=[session_args, decoding], outputs=[session_args]) sampling_temp.change(update_sampling_temp,inputs=[session_args, sampling_temp], outputs=[session_args]) generation_seed.change(update_generation_seed,inputs=[session_args, generation_seed], outputs=[session_args]) n_beams.change(update_n_beams,inputs=[session_args, n_beams], outputs=[session_args]) max_new_tokens.change(update_max_new_tokens,inputs=[session_args, max_new_tokens], outputs=[session_args]) gamma.change(update_gamma,inputs=[session_args, gamma], outputs=[session_args]) delta.change(update_delta,inputs=[session_args, delta], outputs=[session_args]) detection_z_threshold.change(update_detection_z_threshold,inputs=[session_args, detection_z_threshold], outputs=[session_args]) ignore_repeated_bigrams.change(update_ignore_repeated_bigrams,inputs=[session_args, ignore_repeated_bigrams], outputs=[session_args]) normalizers.change(update_normalizers,inputs=[session_args, normalizers], outputs=[session_args]) seed_separately.change(update_seed_separately,inputs=[session_args, seed_separately], outputs=[session_args]) select_green_tokens.change(update_select_green_tokens,inputs=[session_args, select_green_tokens], outputs=[session_args]) # register additional callback on button clicks that updates the shown parameters window generate_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) detect_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) # When the parameters change, display the update and also fire detection, since some detection params dont change the model output. delta.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) gamma.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) gamma.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]) gamma.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) gamma.change(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result,session_args,session_tokenizer,html_detection_input]) detection_z_threshold.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) detection_z_threshold.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]) detection_z_threshold.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) detection_z_threshold.change(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result,session_args,session_tokenizer,html_detection_input]) ignore_repeated_bigrams.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) ignore_repeated_bigrams.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]) ignore_repeated_bigrams.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) ignore_repeated_bigrams.change(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result,session_args,session_tokenizer,html_detection_input]) normalizers.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) normalizers.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]) normalizers.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) normalizers.change(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result,session_args,session_tokenizer,html_detection_input]) select_green_tokens.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters]) select_green_tokens.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]) select_green_tokens.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark]) select_green_tokens.change(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result,session_args,session_tokenizer,html_detection_input]) demo.queue(concurrency_count=3) if args.demo_public: demo.launch(share=True) # exposes app to the internet via randomly generated link else: demo.launch() def main(args): """Run a command line version of the generation and detection operations and optionally launch and serve the gradio demo""" # Initial arg processing and log args.normalizers = (args.normalizers.split(",") if args.normalizers else []) print(args) if not args.skip_model_load: model, tokenizer, device = load_model(args) else: model, tokenizer, device = None, None, None tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path) if args.use_gpu: device = "cuda" if torch.cuda.is_available() else "cpu" else: device = "cpu" # terrapin example input_text = ( "The aardvark is sometimes colloquially called the 'African ant bear',[6] 'anteater' (not to be confused with the South American anteater), or the 'Cape anteater'[6] after the Cape of Good Hope. The name 'aardvark' is Afrikaans (Afrikaans pronunciation: [ˈɑːrtfark]), comes from earlier Afrikaans erdvark[6] and means 'earth pig' or 'ground pig' (aarde: 'earth/ground', vark: 'pig'), because of its burrowing habits.[7][8][9] The name Orycteropus means 'burrowing foot', and the name afer refers to Africa.[10] The name of the aardvark's order, Tubulidentata, comes from the tubule-style teeth.[11]\n\nThe aardvark is not closely related to the pig; rather, it is the sole extant representative of the obscure mammalian order Tubulidentata,[10] in which it is usually considered to form one variable species of the genus Orycteropus, the sole surviving genus in the family Orycteropodidae. The aardvark is" ) args.default_prompt = input_text # Generate and detect, report to stdout if not args.skip_model_load: term_width = 80 print("#"*term_width) print("Prompt:") print(input_text) # a generator that yields (without_watermark, with_watermark) pairs generator_outputs = generate(input_text, args, model=model, device=device, tokenizer=tokenizer) # we need to iterate over it, # but we only want the last output in this case for out in generator_outputs: decoded_output_without_watermark = out[0] decoded_output_with_watermark = out[1] without_watermark_detection_result = detect(decoded_output_without_watermark, args, device=device, tokenizer=tokenizer, return_green_token_mask=False) with_watermark_detection_result = detect(decoded_output_with_watermark, args, device=device, tokenizer=tokenizer, return_green_token_mask=False) print("#"*term_width) print("Output without watermark:") print(decoded_output_without_watermark) print("-"*term_width) print(f"Detection result @ {args.detection_z_threshold}:") pprint(without_watermark_detection_result) print("-"*term_width) print("#"*term_width) print("Output with watermark:") print(decoded_output_with_watermark) print("-"*term_width) print(f"Detection result @ {args.detection_z_threshold}:") pprint(with_watermark_detection_result) print("-"*term_width) # Launch the app to generate and detect interactively (implements the hf space demo) if args.run_gradio: run_gradio(args, model=model, tokenizer=tokenizer, device=device) return if __name__ == "__main__": args = parse_args() print(args) main(args)