Spaces:
Runtime error
Runtime error
File size: 32,208 Bytes
d6b2709 b5b3015 d6b2709 b5b3015 d6b2709 b5b3015 d6b2709 b5b3015 dafc0b4 b5b3015 d6b2709 fa826da d6b2709 b5b3015 6b83b21 d6b2709 a134a9d b5b3015 d6b2709 b5b3015 d6b2709 b5b3015 d6b2709 a134a9d d6b2709 b5b3015 d6b2709 b5b3015 0e6d24f b5b3015 7d29596 b5b3015 7d29596 b5b3015 7d29596 b5b3015 4343565 b5b3015 7d29596 0e6d24f 7d29596 0e6d24f f98590c 0e6d24f f98590c 0e6d24f f98590c 7d29596 0e6d24f f98590c 0e6d24f f98590c 7d29596 0e6d24f b5b3015 0e6d24f 7d29596 f98590c 7d29596 f98590c 7d29596 0e6d24f 7d29596 0e6d24f b5b3015 4343565 b5b3015 7d29596 b5b3015 4343565 b5b3015 7d29596 b5b3015 7d29596 0e6d24f 7d29596 b5b3015 7d29596 b5b3015 7d29596 b5b3015 dafc0b4 4343565 b5b3015 0e6d24f 7d29596 b5b3015 7d29596 0e6d24f b5b3015 4343565 b5b3015 7d29596 b5b3015 0e6d24f 7d29596 0e6d24f 7d29596 a226244 7d29596 a226244 7d29596 a226244 7d29596 a226244 7d29596 a226244 b5b3015 0e6d24f b5b3015 4343565 b5b3015 4343565 b5b3015 a134a9d b5b3015 d6b2709 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 |
# 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):
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():
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):
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):
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):
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 str_format_scores(score_dict, detection_threshold):
# output_str = f"@ z-score threshold={detection_threshold}:\n\n"
# for k,v in score_dict.items():
# if k=='green_fraction':
# output_str+=f"{format_names(k)}={v:.1%}"
# elif k=='confidence':
# output_str+=f"{format_names(k)}={v:.3%}"
# elif isinstance(v, float):
# output_str+=f"{format_names(k)}={v:.3g}"
# else:
# output_str += v
# return output_str
def list_format_scores(score_dict, detection_threshold):
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):
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):
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://github.com/jwkirchenbauer/lm-watermarking)")
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 open source as well as fit on a single commodity GPU. In particular, there are few models above 10B parameters and way fewer trained using both Instruction finetuning or RLHF that are open source that we can use.
Therefore, the model, in both it's un-watermarked (normal) and watermarked state, 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.
"""
)
# 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)
gr.HTML("""
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<br/>
<a href="https://huggingface.co/spaces/tomg-group-umd/lm-watermarking?duplicate=true">
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
<p/>
""")
# Register main generation tab click, outputing generations as well as a the encoded+redecoded+potentially truncated prompt and flag
generate_btn.click(fn=generate_partial, inputs=[prompt,session_args], outputs=[redecoded_input, truncation_warning, output_without_watermark, output_with_watermark,session_args])
# 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])
# Call detection when the outputs (of the generate function) are updated
output_without_watermark.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
output_with_watermark.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
# Register main detection tab click
detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result, session_args])
# State management logic
# update callbacks that change the state dict
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)
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
# registering callbacks for toggling the visibilty of certain parameters
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])
# 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 fire detection, since some detection params dont change the model output.
gamma.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
gamma.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
gamma.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
gamma.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result,session_args])
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], outputs=[without_watermark_detection_result,session_args])
detection_z_threshold.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
detection_z_threshold.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result,session_args])
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], outputs=[without_watermark_detection_result,session_args])
ignore_repeated_bigrams.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
ignore_repeated_bigrams.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result,session_args])
normalizers.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
normalizers.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
normalizers.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
normalizers.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result,session_args])
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], outputs=[without_watermark_detection_result,session_args])
select_green_tokens.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
select_green_tokens.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result,session_args])
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):
# 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
# Generate and detect, report to stdout
if not args.skip_model_load:
# input_text = (
# "The diamondback terrapin or simply terrapin (Malaclemys terrapin) is a "
# "species of turtle native to the brackish coastal tidal marshes of the "
# "Northeastern and southern United States, and in Bermuda.[6] It belongs "
# "to the monotypic genus Malaclemys. It has one of the largest ranges of "
# "all turtles in North America, stretching as far south as the Florida Keys "
# "and as far north as Cape Cod.[7] The name 'terrapin' is derived from the "
# "Algonquian word torope.[8] It applies to Malaclemys terrapin in both "
# "British English and American English. The name originally was used by "
# "early European settlers in North America to describe these brackish-water "
# "turtles that inhabited neither freshwater habitats nor the sea. It retains "
# "this primary meaning in American English.[8] In British English, however, "
# "other semi-aquatic turtle species, such as the red-eared slider, might "
# "also be called terrapins. The common name refers to the diamond pattern "
# "on top of its shell (carapace), but the overall pattern and coloration "
# "vary greatly. The shell is usually wider at the back than in the front, "
# "and from above it appears wedge-shaped. The shell coloring can vary "
# "from brown to grey, and its body color can be grey, brown, yellow, "
# "or white. All have a unique pattern of wiggly, black markings or spots "
# "on their body and head. The diamondback terrapin has large webbed "
# "feet.[9] The species is"
# )
input_text = "In this work, we study watermarking of language model output. A watermark is a hidden pattern in text that is imperceptible to humans, while making the text algorithmically identifiable as synthetic. We propose an efficient watermark that makes synthetic text detectable from short spans of tokens (as few as 25 words), while false-positives (where human text is marked as machine-generated) are statistically improbable. The watermark detection algorithm can be made public, enabling third parties (e.g., social media platforms) to run it themselves, or it can be kept private and run behind an API. We seek a watermark with the following properties:\n"
term_width = 80
print("#"*term_width)
print("Prompt:")
print(input_text)
_, _, decoded_output_without_watermark, decoded_output_with_watermark, _ = generate(input_text,
args,
model=model,
device=device,
tokenizer=tokenizer)
without_watermark_detection_result = detect(decoded_output_without_watermark,
args,
device=device,
tokenizer=tokenizer)
with_watermark_detection_result = detect(decoded_output_with_watermark,
args,
device=device,
tokenizer=tokenizer)
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) |