Spaces:
Sleeping
Sleeping
Commit
·
e2e6e76
1
Parent(s):
60d5b9f
try text-generation 0.3.1 release
Browse files- demo_watermark.py +162 -125
- requirements.txt +2 -2
demo_watermark.py
CHANGED
|
@@ -16,7 +16,6 @@
|
|
| 16 |
|
| 17 |
import os
|
| 18 |
import argparse
|
| 19 |
-
from argparse import Namespace
|
| 20 |
from pprint import pprint
|
| 21 |
from functools import partial
|
| 22 |
|
|
@@ -30,15 +29,25 @@ from transformers import (AutoTokenizer,
|
|
| 30 |
AutoModelForCausalLM,
|
| 31 |
LogitsProcessorList)
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
from watermark_processor import WatermarkLogitsProcessor, WatermarkDetector
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
# FIXME correct lengths for all models
|
| 36 |
API_MODEL_MAP = {
|
| 37 |
-
"bigscience/bloom" : {"max_length":
|
| 38 |
-
"bigscience/bloomz" : {"max_length":
|
| 39 |
-
"google/flan-ul2" : {"max_length":
|
| 40 |
-
"google/flan-t5-xxl" : {"max_length":
|
| 41 |
-
"EleutherAI/gpt-neox-20b" : {"max_length":
|
| 42 |
}
|
| 43 |
|
| 44 |
def str2bool(v):
|
|
@@ -231,35 +240,29 @@ def generate_with_api(prompt, args):
|
|
| 231 |
timeout_msg = "[Model API timeout error. Try reducing the max_new_tokens parameter or the prompt length.]"
|
| 232 |
try:
|
| 233 |
generation_params["watermark"] = False
|
| 234 |
-
|
| 235 |
-
output_text_without_watermark = output.generated_text
|
| 236 |
except ReadTimeout as e:
|
| 237 |
print(e)
|
| 238 |
-
|
| 239 |
try:
|
| 240 |
generation_params["watermark"] = True
|
| 241 |
-
|
| 242 |
-
output_text_with_watermark = output.generated_text
|
| 243 |
except ReadTimeout as e:
|
| 244 |
print(e)
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
return (output_text_without_watermark,
|
| 248 |
-
output_text_with_watermark)
|
| 249 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 250 |
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
and generate watermarked text by passing it to the generate method of the model
|
| 254 |
-
as a logits processor. """
|
| 255 |
-
|
| 256 |
-
print(f"Generating with {args}")
|
| 257 |
|
| 258 |
# This applies to both the local and API model scenarios
|
| 259 |
-
if args.
|
| 260 |
-
|
| 261 |
-
elif args.model_name_or_path in API_MODEL_MAP:
|
| 262 |
-
args.prompt_max_length = API_MODEL_MAP[args.model_name_or_path]["max_length"]-args.max_new_tokens
|
| 263 |
elif hasattr(model.config,"max_position_embedding"):
|
| 264 |
args.prompt_max_length = model.config.max_position_embeddings-args.max_new_tokens
|
| 265 |
else:
|
|
@@ -269,69 +272,77 @@ def generate(prompt, args, tokenizer, model=None, device=None):
|
|
| 269 |
truncation_warning = True if tokd_input["input_ids"].shape[-1] == args.prompt_max_length else False
|
| 270 |
redecoded_input = tokenizer.batch_decode(tokd_input["input_ids"], skip_special_tokens=True)[0]
|
| 271 |
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
decoded_output_with_watermark = api_outputs[1]
|
| 276 |
-
return (redecoded_input,
|
| 277 |
-
int(truncation_warning),
|
| 278 |
-
decoded_output_without_watermark,
|
| 279 |
-
decoded_output_with_watermark,
|
| 280 |
-
args,
|
| 281 |
-
tokenizer)
|
| 282 |
-
|
| 283 |
|
| 284 |
-
watermark_processor = WatermarkLogitsProcessor(vocab=list(tokenizer.get_vocab().values()),
|
| 285 |
-
gamma=args.gamma,
|
| 286 |
-
delta=args.delta,
|
| 287 |
-
seeding_scheme=args.seeding_scheme,
|
| 288 |
-
select_green_tokens=args.select_green_tokens)
|
| 289 |
|
| 290 |
-
gen_kwargs = dict(max_new_tokens=args.max_new_tokens)
|
| 291 |
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
temperature=args.sampling_temp
|
| 297 |
-
))
|
| 298 |
-
else:
|
| 299 |
-
gen_kwargs.update(dict(
|
| 300 |
-
num_beams=args.n_beams
|
| 301 |
-
))
|
| 302 |
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
**gen_kwargs
|
| 306 |
-
)
|
| 307 |
-
generate_with_watermark = partial(
|
| 308 |
-
model.generate,
|
| 309 |
-
logits_processor=LogitsProcessorList([watermark_processor]),
|
| 310 |
-
**gen_kwargs
|
| 311 |
-
)
|
| 312 |
|
| 313 |
-
|
| 314 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 315 |
|
| 316 |
-
# optional to seed before second generation, but will not be the same again generally, unless delta==0.0, no-op watermark
|
| 317 |
-
if args.seed_separately:
|
| 318 |
torch.manual_seed(args.generation_seed)
|
| 319 |
-
|
| 320 |
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
output_with_watermark =
|
| 325 |
|
| 326 |
-
|
| 327 |
-
|
|
|
|
|
|
|
| 328 |
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
|
|
|
|
|
|
|
|
|
| 335 |
|
| 336 |
|
| 337 |
def format_names(s):
|
|
@@ -348,7 +359,6 @@ def format_names(s):
|
|
| 348 |
def list_format_scores(score_dict, detection_threshold):
|
| 349 |
"""Format the detection metrics into a gradio dataframe input format"""
|
| 350 |
lst_2d = []
|
| 351 |
-
# lst_2d.append(["z-score threshold", f"{detection_threshold}"])
|
| 352 |
for k,v in score_dict.items():
|
| 353 |
if k=='green_fraction':
|
| 354 |
lst_2d.append([format_names(k), f"{v:.1%}"])
|
|
@@ -366,9 +376,10 @@ def list_format_scores(score_dict, detection_threshold):
|
|
| 366 |
lst_2d.insert(-1,["z-score Threshold", f"{detection_threshold}"])
|
| 367 |
return lst_2d
|
| 368 |
|
| 369 |
-
def detect(input_text, args, tokenizer, device=None):
|
| 370 |
"""Instantiate the WatermarkDetection object and call detect on
|
| 371 |
the input text returning the scores and outcome of the test"""
|
|
|
|
| 372 |
print(f"Detecting with {args}")
|
| 373 |
print(f"Detection Tokenizer: {type(tokenizer)}")
|
| 374 |
|
|
@@ -381,14 +392,14 @@ def detect(input_text, args, tokenizer, device=None):
|
|
| 381 |
normalizers=args.normalizers,
|
| 382 |
ignore_repeated_bigrams=args.ignore_repeated_bigrams,
|
| 383 |
select_green_tokens=args.select_green_tokens)
|
| 384 |
-
# if len(input_text)-1 > watermark_detector.min_prefix_len:
|
| 385 |
error = False
|
|
|
|
| 386 |
if input_text == "":
|
| 387 |
error = True
|
| 388 |
else:
|
| 389 |
-
try:
|
| 390 |
-
score_dict = watermark_detector.detect(input_text)
|
| 391 |
-
|
| 392 |
output = list_format_scores(score_dict, watermark_detector.z_threshold)
|
| 393 |
except ValueError as e:
|
| 394 |
print(e)
|
|
@@ -396,16 +407,41 @@ def detect(input_text, args, tokenizer, device=None):
|
|
| 396 |
if error:
|
| 397 |
output = [["Error","string too short to compute metrics"]]
|
| 398 |
output += [["",""] for _ in range(6)]
|
| 399 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 400 |
|
| 401 |
def run_gradio(args, model=None, device=None, tokenizer=None):
|
| 402 |
"""Define and launch the gradio demo interface"""
|
| 403 |
-
|
| 404 |
-
# detect_partial = partial(detect, device=device, tokenizer=tokenizer)
|
| 405 |
generate_partial = partial(generate, model=model, device=device)
|
| 406 |
detect_partial = partial(detect, device=device)
|
| 407 |
|
| 408 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 409 |
# Top section, greeting and instructions
|
| 410 |
with gr.Row():
|
| 411 |
with gr.Column(scale=9):
|
|
@@ -422,7 +458,6 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 422 |
[](https://github.com/jwkirchenbauer/lm-watermarking)
|
| 423 |
"""
|
| 424 |
)
|
| 425 |
-
# gr.Markdown(f"Language model: {args.model_name_or_path} {'(float16 mode)' if args.load_fp16 else ''}")
|
| 426 |
# if model_name_or_path at startup not one of the API models then add to dropdown
|
| 427 |
all_models = sorted(list(set(list(API_MODEL_MAP.keys())+[args.model_name_or_path])))
|
| 428 |
model_selector = gr.Dropdown(
|
|
@@ -475,8 +510,10 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 475 |
"""
|
| 476 |
**[Generate & Detect]**: The first tab shows that the watermark can be embedded with
|
| 477 |
negligible impact on text quality. You can try any prompt and compare the quality of
|
| 478 |
-
normal text (*Output Without Watermark*) to the watermarked text (*Output With Watermark*) below it.
|
| 479 |
-
|
|
|
|
|
|
|
| 480 |
Detection is very efficient and does not use the language model or its parameters.
|
| 481 |
|
| 482 |
**[Detector Only]**: You can also copy-paste the watermarked text (or any other text)
|
|
@@ -495,7 +532,6 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 495 |
"""
|
| 496 |
)
|
| 497 |
|
| 498 |
-
|
| 499 |
with gr.Tab("Generate & Detect"):
|
| 500 |
|
| 501 |
with gr.Row():
|
|
@@ -504,15 +540,19 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 504 |
generate_btn = gr.Button("Generate")
|
| 505 |
with gr.Row():
|
| 506 |
with gr.Column(scale=2):
|
| 507 |
-
|
|
|
|
|
|
|
|
|
|
| 508 |
with gr.Column(scale=1):
|
| 509 |
-
# without_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14)
|
| 510 |
without_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
| 511 |
with gr.Row():
|
| 512 |
with gr.Column(scale=2):
|
| 513 |
-
|
|
|
|
|
|
|
|
|
|
| 514 |
with gr.Column(scale=1):
|
| 515 |
-
# with_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14)
|
| 516 |
with_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"],interactive=False,row_count=7,col_count=2)
|
| 517 |
|
| 518 |
redecoded_input = gr.Textbox(visible=False)
|
|
@@ -528,7 +568,6 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 528 |
with gr.Column(scale=2):
|
| 529 |
detection_input = gr.Textbox(label="Text to Analyze", interactive=True,lines=14,max_lines=14)
|
| 530 |
with gr.Column(scale=1):
|
| 531 |
-
# detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=14,max_lines=14)
|
| 532 |
detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
| 533 |
with gr.Row():
|
| 534 |
detect_btn = gr.Button("Detect")
|
|
@@ -562,7 +601,6 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 562 |
ignore_repeated_bigrams = gr.Checkbox(label="Ignore Bigram Repeats")
|
| 563 |
with gr.Row():
|
| 564 |
normalizers = gr.CheckboxGroup(label="Normalizations", choices=["unicode", "homoglyphs", "truecase"], value=args.normalizers)
|
| 565 |
-
# with gr.Accordion("Actual submitted parameters:",open=False):
|
| 566 |
with gr.Row():
|
| 567 |
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._")
|
| 568 |
with gr.Row():
|
|
@@ -657,15 +695,14 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 657 |
<p/>
|
| 658 |
""")
|
| 659 |
|
| 660 |
-
# Register main generation tab click, outputing generations as well as a the encoded+redecoded+potentially truncated prompt and flag
|
| 661 |
-
generate_btn.click(fn=
|
|
|
|
|
|
|
|
|
|
| 662 |
# Show truncated version of prompt if truncation occurred
|
| 663 |
redecoded_input.change(fn=truncate_prompt, inputs=[redecoded_input,truncation_warning,prompt,session_args], outputs=[prompt,session_args])
|
| 664 |
-
# Call detection when the outputs (of the generate function) are updated
|
| 665 |
-
output_without_watermark.change(fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer])
|
| 666 |
-
output_with_watermark.change(fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer])
|
| 667 |
# Register main detection tab click
|
| 668 |
-
# detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result, session_args,session_tokenizer])
|
| 669 |
detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result, session_args,session_tokenizer], api_name="detection")
|
| 670 |
|
| 671 |
# State management logic
|
|
@@ -727,7 +764,11 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
| 727 |
def update_normalizers(session_state, value): session_state.normalizers = value; return session_state
|
| 728 |
def update_seed_separately(session_state, value): session_state.seed_separately = value; return session_state
|
| 729 |
def update_select_green_tokens(session_state, value): session_state.select_green_tokens = value; return session_state
|
| 730 |
-
def update_tokenizer(model_name_or_path):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 731 |
# registering callbacks for toggling the visibilty of certain parameters based on the values of others
|
| 732 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[sampling_temp])
|
| 733 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[generation_seed])
|
|
@@ -829,19 +870,6 @@ def main(args):
|
|
| 829 |
"on their body and head. The diamondback terrapin has large webbed "
|
| 830 |
"feet.[9] The species is"
|
| 831 |
)
|
| 832 |
-
|
| 833 |
-
# teaser example
|
| 834 |
-
# input_text = (
|
| 835 |
-
# "In this work, we study watermarking of language model output. "
|
| 836 |
-
# "A watermark is a hidden pattern in text that is imperceptible to humans, "
|
| 837 |
-
# "while making the text algorithmically identifiable as synthetic. "
|
| 838 |
-
# "We propose an efficient watermark that makes synthetic text detectable "
|
| 839 |
-
# "from short spans of tokens (as few as 25 words), while false-positives "
|
| 840 |
-
# "(where human text is marked as machine-generated) are statistically improbable. "
|
| 841 |
-
# "The watermark detection algorithm can be made public, enabling third parties "
|
| 842 |
-
# "(e.g., social media platforms) to run it themselves, or it can be kept private "
|
| 843 |
-
# "and run behind an API. We seek a watermark with the following properties:\n"
|
| 844 |
-
# )
|
| 845 |
|
| 846 |
args.default_prompt = input_text
|
| 847 |
|
|
@@ -854,19 +882,28 @@ def main(args):
|
|
| 854 |
print("Prompt:")
|
| 855 |
print(input_text)
|
| 856 |
|
| 857 |
-
|
| 858 |
-
|
| 859 |
-
|
| 860 |
-
|
| 861 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 862 |
without_watermark_detection_result = detect(decoded_output_without_watermark,
|
| 863 |
args,
|
| 864 |
device=device,
|
| 865 |
-
tokenizer=tokenizer
|
|
|
|
| 866 |
with_watermark_detection_result = detect(decoded_output_with_watermark,
|
| 867 |
args,
|
| 868 |
device=device,
|
| 869 |
-
tokenizer=tokenizer
|
|
|
|
| 870 |
|
| 871 |
print("#"*term_width)
|
| 872 |
print("Output without watermark:")
|
|
|
|
| 16 |
|
| 17 |
import os
|
| 18 |
import argparse
|
|
|
|
| 19 |
from pprint import pprint
|
| 20 |
from functools import partial
|
| 21 |
|
|
|
|
| 29 |
AutoModelForCausalLM,
|
| 30 |
LogitsProcessorList)
|
| 31 |
|
| 32 |
+
# from local_tokenizers.tokenization_llama import LLaMATokenizer
|
| 33 |
+
|
| 34 |
+
from transformers import GPT2TokenizerFast
|
| 35 |
+
OPT_TOKENIZER = GPT2TokenizerFast
|
| 36 |
+
|
| 37 |
from watermark_processor import WatermarkLogitsProcessor, WatermarkDetector
|
| 38 |
|
| 39 |
+
|
| 40 |
+
# ALPACA_MODEL_NAME = "alpaca"
|
| 41 |
+
# ALPACA_MODEL_TOKENIZER = LLaMATokenizer
|
| 42 |
+
# ALPACA_TOKENIZER_PATH = "/cmlscratch/jkirchen/llama"
|
| 43 |
+
|
| 44 |
# FIXME correct lengths for all models
|
| 45 |
API_MODEL_MAP = {
|
| 46 |
+
"bigscience/bloom" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0},
|
| 47 |
+
"bigscience/bloomz" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0},
|
| 48 |
+
"google/flan-ul2" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0},
|
| 49 |
+
"google/flan-t5-xxl" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0},
|
| 50 |
+
"EleutherAI/gpt-neox-20b" : {"max_length": 1000, "gamma": 0.5, "delta": 2.0},
|
| 51 |
}
|
| 52 |
|
| 53 |
def str2bool(v):
|
|
|
|
| 240 |
timeout_msg = "[Model API timeout error. Try reducing the max_new_tokens parameter or the prompt length.]"
|
| 241 |
try:
|
| 242 |
generation_params["watermark"] = False
|
| 243 |
+
without_watermark_iterator = client.generate_stream(prompt, **generation_params)
|
|
|
|
| 244 |
except ReadTimeout as e:
|
| 245 |
print(e)
|
| 246 |
+
without_watermark_iterator = (char for char in timeout_msg)
|
| 247 |
try:
|
| 248 |
generation_params["watermark"] = True
|
| 249 |
+
with_watermark_iterator = client.generate_stream(prompt, **generation_params)
|
|
|
|
| 250 |
except ReadTimeout as e:
|
| 251 |
print(e)
|
| 252 |
+
with_watermark_iterator = (char for char in timeout_msg)
|
|
|
|
|
|
|
|
|
|
| 253 |
|
| 254 |
+
all_without_words, all_with_words = "", ""
|
| 255 |
+
for without_word, with_word in zip(without_watermark_iterator, with_watermark_iterator):
|
| 256 |
+
all_without_words += without_word.token.text
|
| 257 |
+
all_with_words += with_word.token.text
|
| 258 |
+
yield all_without_words, all_with_words
|
| 259 |
|
| 260 |
+
|
| 261 |
+
def check_prompt(prompt, args, tokenizer, model=None, device=None):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 262 |
|
| 263 |
# This applies to both the local and API model scenarios
|
| 264 |
+
if args.model_name_or_path in API_MODEL_MAP:
|
| 265 |
+
args.prompt_max_length = API_MODEL_MAP[args.model_name_or_path]["max_length"]
|
|
|
|
|
|
|
| 266 |
elif hasattr(model.config,"max_position_embedding"):
|
| 267 |
args.prompt_max_length = model.config.max_position_embeddings-args.max_new_tokens
|
| 268 |
else:
|
|
|
|
| 272 |
truncation_warning = True if tokd_input["input_ids"].shape[-1] == args.prompt_max_length else False
|
| 273 |
redecoded_input = tokenizer.batch_decode(tokd_input["input_ids"], skip_special_tokens=True)[0]
|
| 274 |
|
| 275 |
+
return (redecoded_input,
|
| 276 |
+
int(truncation_warning),
|
| 277 |
+
args)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 278 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 279 |
|
|
|
|
| 280 |
|
| 281 |
+
def generate(prompt, args, tokenizer, model=None, device=None):
|
| 282 |
+
"""Instatiate the WatermarkLogitsProcessor according to the watermark parameters
|
| 283 |
+
and generate watermarked text by passing it to the generate method of the model
|
| 284 |
+
as a logits processor. """
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 285 |
|
| 286 |
+
print(f"Generating with {args}")
|
| 287 |
+
print(f"Prompt: {prompt}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 288 |
|
| 289 |
+
if args.model_name_or_path in API_MODEL_MAP:
|
| 290 |
+
api_outputs = generate_with_api(prompt, args)
|
| 291 |
+
yield from api_outputs
|
| 292 |
+
else:
|
| 293 |
+
tokd_input = tokenizer(prompt, return_tensors="pt", add_special_tokens=True, truncation=True, max_length=args.prompt_max_length).to(device)
|
| 294 |
+
|
| 295 |
+
watermark_processor = WatermarkLogitsProcessor(vocab=list(tokenizer.get_vocab().values()),
|
| 296 |
+
gamma=args.gamma,
|
| 297 |
+
delta=args.delta,
|
| 298 |
+
seeding_scheme=args.seeding_scheme,
|
| 299 |
+
select_green_tokens=args.select_green_tokens)
|
| 300 |
+
|
| 301 |
+
gen_kwargs = dict(max_new_tokens=args.max_new_tokens)
|
| 302 |
+
|
| 303 |
+
if args.use_sampling:
|
| 304 |
+
gen_kwargs.update(dict(
|
| 305 |
+
do_sample=True,
|
| 306 |
+
top_k=0,
|
| 307 |
+
temperature=args.sampling_temp
|
| 308 |
+
))
|
| 309 |
+
else:
|
| 310 |
+
gen_kwargs.update(dict(
|
| 311 |
+
num_beams=args.n_beams
|
| 312 |
+
))
|
| 313 |
+
|
| 314 |
+
generate_without_watermark = partial(
|
| 315 |
+
model.generate,
|
| 316 |
+
**gen_kwargs
|
| 317 |
+
)
|
| 318 |
+
generate_with_watermark = partial(
|
| 319 |
+
model.generate,
|
| 320 |
+
logits_processor=LogitsProcessorList([watermark_processor]),
|
| 321 |
+
**gen_kwargs
|
| 322 |
+
)
|
| 323 |
|
|
|
|
|
|
|
| 324 |
torch.manual_seed(args.generation_seed)
|
| 325 |
+
output_without_watermark = generate_without_watermark(**tokd_input)
|
| 326 |
|
| 327 |
+
# optional to seed before second generation, but will not be the same again generally, unless delta==0.0, no-op watermark
|
| 328 |
+
if args.seed_separately:
|
| 329 |
+
torch.manual_seed(args.generation_seed)
|
| 330 |
+
output_with_watermark = generate_with_watermark(**tokd_input)
|
| 331 |
|
| 332 |
+
if args.is_decoder_only_model:
|
| 333 |
+
# need to isolate the newly generated tokens
|
| 334 |
+
output_without_watermark = output_without_watermark[:,tokd_input["input_ids"].shape[-1]:]
|
| 335 |
+
output_with_watermark = output_with_watermark[:,tokd_input["input_ids"].shape[-1]:]
|
| 336 |
|
| 337 |
+
decoded_output_without_watermark = tokenizer.batch_decode(output_without_watermark, skip_special_tokens=True)[0]
|
| 338 |
+
decoded_output_with_watermark = tokenizer.batch_decode(output_with_watermark, skip_special_tokens=True)[0]
|
| 339 |
+
|
| 340 |
+
# mocking the API outputs in a whitespace split generator style
|
| 341 |
+
all_without_words, all_with_words = "", ""
|
| 342 |
+
for without_word, with_word in zip(decoded_output_without_watermark.split(), decoded_output_with_watermark.split()):
|
| 343 |
+
all_without_words += without_word + " "
|
| 344 |
+
all_with_words += with_word + " "
|
| 345 |
+
yield all_without_words, all_with_words
|
| 346 |
|
| 347 |
|
| 348 |
def format_names(s):
|
|
|
|
| 359 |
def list_format_scores(score_dict, detection_threshold):
|
| 360 |
"""Format the detection metrics into a gradio dataframe input format"""
|
| 361 |
lst_2d = []
|
|
|
|
| 362 |
for k,v in score_dict.items():
|
| 363 |
if k=='green_fraction':
|
| 364 |
lst_2d.append([format_names(k), f"{v:.1%}"])
|
|
|
|
| 376 |
lst_2d.insert(-1,["z-score Threshold", f"{detection_threshold}"])
|
| 377 |
return lst_2d
|
| 378 |
|
| 379 |
+
def detect(input_text, args, tokenizer, device=None, return_green_token_mask=True):
|
| 380 |
"""Instantiate the WatermarkDetection object and call detect on
|
| 381 |
the input text returning the scores and outcome of the test"""
|
| 382 |
+
|
| 383 |
print(f"Detecting with {args}")
|
| 384 |
print(f"Detection Tokenizer: {type(tokenizer)}")
|
| 385 |
|
|
|
|
| 392 |
normalizers=args.normalizers,
|
| 393 |
ignore_repeated_bigrams=args.ignore_repeated_bigrams,
|
| 394 |
select_green_tokens=args.select_green_tokens)
|
|
|
|
| 395 |
error = False
|
| 396 |
+
green_token_mask = None
|
| 397 |
if input_text == "":
|
| 398 |
error = True
|
| 399 |
else:
|
| 400 |
+
try:
|
| 401 |
+
score_dict = watermark_detector.detect(input_text, return_green_token_mask=return_green_token_mask)
|
| 402 |
+
green_token_mask = score_dict.pop("green_token_mask", None)
|
| 403 |
output = list_format_scores(score_dict, watermark_detector.z_threshold)
|
| 404 |
except ValueError as e:
|
| 405 |
print(e)
|
|
|
|
| 407 |
if error:
|
| 408 |
output = [["Error","string too short to compute metrics"]]
|
| 409 |
output += [["",""] for _ in range(6)]
|
| 410 |
+
|
| 411 |
+
html_output = ""
|
| 412 |
+
if green_token_mask is not None:
|
| 413 |
+
# hack bc we need a fast tokenizer with charspan support
|
| 414 |
+
if "opt" in args.model_name_or_path:
|
| 415 |
+
tokenizer = OPT_TOKENIZER.from_pretrained(args.model_name_or_path)
|
| 416 |
+
|
| 417 |
+
tokens = tokenizer(input_text)
|
| 418 |
+
if tokens["input_ids"][0] == tokenizer.bos_token_id:
|
| 419 |
+
tokens["input_ids"] = tokens["input_ids"][1:] # ignore attention mask
|
| 420 |
+
skip = watermark_detector.min_prefix_len
|
| 421 |
+
charspans = [tokens.token_to_chars(i) for i in range(skip,len(tokens["input_ids"]))]
|
| 422 |
+
charspans = [cs for cs in charspans if cs is not None] # remove the special token spans
|
| 423 |
+
|
| 424 |
+
if len(charspans) != len(green_token_mask): breakpoint()
|
| 425 |
+
assert len(charspans) == len(green_token_mask)
|
| 426 |
+
|
| 427 |
+
tags = [(f'<span class="green">{input_text[cs.start:cs.end]}</span>' if m else f'<span class="red">{input_text[cs.start:cs.end]}</span>') for cs, m in zip(charspans, green_token_mask)]
|
| 428 |
+
html_output = f'<p>{" ".join(tags)}</p>'
|
| 429 |
+
|
| 430 |
+
return output, args, tokenizer, html_output
|
| 431 |
|
| 432 |
def run_gradio(args, model=None, device=None, tokenizer=None):
|
| 433 |
"""Define and launch the gradio demo interface"""
|
| 434 |
+
check_prompt_partial = partial(check_prompt, model=model, device=device)
|
|
|
|
| 435 |
generate_partial = partial(generate, model=model, device=device)
|
| 436 |
detect_partial = partial(detect, device=device)
|
| 437 |
|
| 438 |
+
|
| 439 |
+
css = """
|
| 440 |
+
.green { color: black!important;line-height:1.9em; padding: 0.2em 0.2em; background: #ccffcc; border-radius:0.5rem;}
|
| 441 |
+
.red { color: black!important;line-height:1.9em; padding: 0.2em 0.2em; background: #ffad99; border-radius:0.5rem;}
|
| 442 |
+
"""
|
| 443 |
+
|
| 444 |
+
with gr.Blocks(css=css) as demo:
|
| 445 |
# Top section, greeting and instructions
|
| 446 |
with gr.Row():
|
| 447 |
with gr.Column(scale=9):
|
|
|
|
| 458 |
[](https://github.com/jwkirchenbauer/lm-watermarking)
|
| 459 |
"""
|
| 460 |
)
|
|
|
|
| 461 |
# if model_name_or_path at startup not one of the API models then add to dropdown
|
| 462 |
all_models = sorted(list(set(list(API_MODEL_MAP.keys())+[args.model_name_or_path])))
|
| 463 |
model_selector = gr.Dropdown(
|
|
|
|
| 510 |
"""
|
| 511 |
**[Generate & Detect]**: The first tab shows that the watermark can be embedded with
|
| 512 |
negligible impact on text quality. You can try any prompt and compare the quality of
|
| 513 |
+
normal text (*Output Without Watermark*) to the watermarked text (*Output With Watermark*) below it.
|
| 514 |
+
You can also "see" the watermark by looking at the **Highlighted** tab where the tokens are
|
| 515 |
+
colored green or red depending on which list they are in.
|
| 516 |
+
Metrics on the right show that the watermark can be reliably detected given a reasonably small number of tokens (25-50).
|
| 517 |
Detection is very efficient and does not use the language model or its parameters.
|
| 518 |
|
| 519 |
**[Detector Only]**: You can also copy-paste the watermarked text (or any other text)
|
|
|
|
| 532 |
"""
|
| 533 |
)
|
| 534 |
|
|
|
|
| 535 |
with gr.Tab("Generate & Detect"):
|
| 536 |
|
| 537 |
with gr.Row():
|
|
|
|
| 540 |
generate_btn = gr.Button("Generate")
|
| 541 |
with gr.Row():
|
| 542 |
with gr.Column(scale=2):
|
| 543 |
+
with gr.Tab("Output Without Watermark (Raw Text)"):
|
| 544 |
+
output_without_watermark = gr.Textbox(interactive=False,lines=14,max_lines=14)
|
| 545 |
+
with gr.Tab("Highlighted"):
|
| 546 |
+
html_without_watermark = gr.HTML(elem_id="html-without-watermark")
|
| 547 |
with gr.Column(scale=1):
|
|
|
|
| 548 |
without_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
| 549 |
with gr.Row():
|
| 550 |
with gr.Column(scale=2):
|
| 551 |
+
with gr.Tab("Output With Watermark (Raw Text)"):
|
| 552 |
+
output_with_watermark = gr.Textbox(interactive=False,lines=14,max_lines=14)
|
| 553 |
+
with gr.Tab("Highlighted"):
|
| 554 |
+
html_with_watermark = gr.HTML(elem_id="html-with-watermark")
|
| 555 |
with gr.Column(scale=1):
|
|
|
|
| 556 |
with_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"],interactive=False,row_count=7,col_count=2)
|
| 557 |
|
| 558 |
redecoded_input = gr.Textbox(visible=False)
|
|
|
|
| 568 |
with gr.Column(scale=2):
|
| 569 |
detection_input = gr.Textbox(label="Text to Analyze", interactive=True,lines=14,max_lines=14)
|
| 570 |
with gr.Column(scale=1):
|
|
|
|
| 571 |
detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
| 572 |
with gr.Row():
|
| 573 |
detect_btn = gr.Button("Detect")
|
|
|
|
| 601 |
ignore_repeated_bigrams = gr.Checkbox(label="Ignore Bigram Repeats")
|
| 602 |
with gr.Row():
|
| 603 |
normalizers = gr.CheckboxGroup(label="Normalizations", choices=["unicode", "homoglyphs", "truecase"], value=args.normalizers)
|
|
|
|
| 604 |
with gr.Row():
|
| 605 |
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._")
|
| 606 |
with gr.Row():
|
|
|
|
| 695 |
<p/>
|
| 696 |
""")
|
| 697 |
|
| 698 |
+
# Register main generation tab click, outputing generations as well as a the encoded+redecoded+potentially truncated prompt and flag, then call detection
|
| 699 |
+
generate_btn.click(fn=check_prompt_partial, inputs=[prompt,session_args,session_tokenizer], outputs=[redecoded_input, truncation_warning, session_args]).success(
|
| 700 |
+
fn=generate_partial, inputs=[redecoded_input,session_args,session_tokenizer], outputs=[output_without_watermark, output_with_watermark]).success(
|
| 701 |
+
fn=detect_partial, inputs=[output_without_watermark,session_args,session_tokenizer], outputs=[without_watermark_detection_result,session_args,session_tokenizer,html_without_watermark]).success(
|
| 702 |
+
fn=detect_partial, inputs=[output_with_watermark,session_args,session_tokenizer], outputs=[with_watermark_detection_result,session_args,session_tokenizer,html_with_watermark])
|
| 703 |
# Show truncated version of prompt if truncation occurred
|
| 704 |
redecoded_input.change(fn=truncate_prompt, inputs=[redecoded_input,truncation_warning,prompt,session_args], outputs=[prompt,session_args])
|
|
|
|
|
|
|
|
|
|
| 705 |
# Register main detection tab click
|
|
|
|
| 706 |
detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args,session_tokenizer], outputs=[detection_result, session_args,session_tokenizer], api_name="detection")
|
| 707 |
|
| 708 |
# State management logic
|
|
|
|
| 764 |
def update_normalizers(session_state, value): session_state.normalizers = value; return session_state
|
| 765 |
def update_seed_separately(session_state, value): session_state.seed_separately = value; return session_state
|
| 766 |
def update_select_green_tokens(session_state, value): session_state.select_green_tokens = value; return session_state
|
| 767 |
+
def update_tokenizer(model_name_or_path):
|
| 768 |
+
# if model_name_or_path == ALPACA_MODEL_NAME:
|
| 769 |
+
# return ALPACA_MODEL_TOKENIZER.from_pretrained(ALPACA_TOKENIZER_PATH)
|
| 770 |
+
# else:
|
| 771 |
+
return AutoTokenizer.from_pretrained(model_name_or_path)
|
| 772 |
# registering callbacks for toggling the visibilty of certain parameters based on the values of others
|
| 773 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[sampling_temp])
|
| 774 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[generation_seed])
|
|
|
|
| 870 |
"on their body and head. The diamondback terrapin has large webbed "
|
| 871 |
"feet.[9] The species is"
|
| 872 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 873 |
|
| 874 |
args.default_prompt = input_text
|
| 875 |
|
|
|
|
| 882 |
print("Prompt:")
|
| 883 |
print(input_text)
|
| 884 |
|
| 885 |
+
# a generator that yields (without_watermark, with_watermark) pairs
|
| 886 |
+
generator_outputs = generate(input_text,
|
| 887 |
+
args,
|
| 888 |
+
model=model,
|
| 889 |
+
device=device,
|
| 890 |
+
tokenizer=tokenizer)
|
| 891 |
+
# we need to iterate over it,
|
| 892 |
+
# but we only want the last output in this case
|
| 893 |
+
for out in generator_outputs:
|
| 894 |
+
decoded_output_without_watermark = out[0]
|
| 895 |
+
decoded_output_with_watermark = out[1]
|
| 896 |
+
|
| 897 |
without_watermark_detection_result = detect(decoded_output_without_watermark,
|
| 898 |
args,
|
| 899 |
device=device,
|
| 900 |
+
tokenizer=tokenizer,
|
| 901 |
+
return_green_token_mask=False)
|
| 902 |
with_watermark_detection_result = detect(decoded_output_with_watermark,
|
| 903 |
args,
|
| 904 |
device=device,
|
| 905 |
+
tokenizer=tokenizer,
|
| 906 |
+
return_green_token_mask=False)
|
| 907 |
|
| 908 |
print("#"*term_width)
|
| 909 |
print("Output without watermark:")
|
requirements.txt
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
spacy
|
| 2 |
-
gradio
|
| 3 |
nltk
|
| 4 |
scipy
|
| 5 |
torch
|
| 6 |
transformers
|
| 7 |
tokenizers
|
| 8 |
accelerate
|
| 9 |
-
text-generation>=0.3.
|
|
|
|
| 1 |
spacy
|
| 2 |
+
gradio>=3.21.0
|
| 3 |
nltk
|
| 4 |
scipy
|
| 5 |
torch
|
| 6 |
transformers
|
| 7 |
tokenizers
|
| 8 |
accelerate
|
| 9 |
+
text-generation>=0.3.1
|