Spaces:
Sleeping
Sleeping
jwkirchenbauer
commited on
Commit
•
7d29596
1
Parent(s):
dafc0b4
more polished interface
Browse files- app.py +9 -3
- demo_watermark.py +124 -55
- homoglyph_data/__init__.py +40 -0
- homoglyph_data/categories.json +0 -0
- homoglyph_data/confusables_sept2022.json +0 -0
- homoglyph_data/languages.json +34 -0
- homoglyphs.py +11 -14
- requirements.txt +0 -1
- watermark_processor.py +2 -2
app.py
CHANGED
@@ -19,9 +19,14 @@ args = Namespace()
|
|
19 |
|
20 |
arg_dict = {
|
21 |
'run_gradio': True,
|
22 |
-
'demo_public': False,
|
23 |
-
|
24 |
-
'model_name_or_path': 'facebook/opt-
|
|
|
|
|
|
|
|
|
|
|
25 |
'prompt_max_length': None,
|
26 |
'max_new_tokens': 200,
|
27 |
'generation_seed': 123,
|
@@ -36,6 +41,7 @@ arg_dict = {
|
|
36 |
'ignore_repeated_bigrams': False,
|
37 |
'detection_z_threshold': 4.0,
|
38 |
'select_green_tokens': True,
|
|
|
39 |
'skip_model_load': False,
|
40 |
'seed_separately': True,
|
41 |
}
|
|
|
19 |
|
20 |
arg_dict = {
|
21 |
'run_gradio': True,
|
22 |
+
# 'demo_public': False,
|
23 |
+
'demo_public': True,
|
24 |
+
'model_name_or_path': 'facebook/opt-125m',
|
25 |
+
# 'model_name_or_path': 'facebook/opt-1.3b',
|
26 |
+
# 'model_name_or_path': 'facebook/opt-2.7b',
|
27 |
+
# 'model_name_or_path': 'facebook/opt-6.7b',
|
28 |
+
# 'model_name_or_path': 'facebook/opt-13b',
|
29 |
+
# 'model_name_or_path': 'facebook/opt-30b',
|
30 |
'prompt_max_length': None,
|
31 |
'max_new_tokens': 200,
|
32 |
'generation_seed': 123,
|
|
|
41 |
'ignore_repeated_bigrams': False,
|
42 |
'detection_z_threshold': 4.0,
|
43 |
'select_green_tokens': True,
|
44 |
+
# 'skip_model_load': True,
|
45 |
'skip_model_load': False,
|
46 |
'seed_separately': True,
|
47 |
}
|
demo_watermark.py
CHANGED
@@ -250,6 +250,41 @@ def generate(prompt, args, model=None, device=None, tokenizer=None):
|
|
250 |
args)
|
251 |
# decoded_output_with_watermark)
|
252 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
def detect(input_text, args, device=None, tokenizer=None):
|
254 |
watermark_detector = WatermarkDetector(vocab=list(tokenizer.get_vocab().values()),
|
255 |
gamma=args.gamma,
|
@@ -262,11 +297,13 @@ def detect(input_text, args, device=None, tokenizer=None):
|
|
262 |
select_green_tokens=args.select_green_tokens)
|
263 |
if len(input_text)-1 > watermark_detector.min_prefix_len:
|
264 |
score_dict = watermark_detector.detect(input_text)
|
265 |
-
|
266 |
-
|
267 |
else:
|
268 |
-
|
269 |
-
|
|
|
|
|
270 |
|
271 |
def run_gradio(args, model=None, device=None, tokenizer=None):
|
272 |
|
@@ -276,33 +313,41 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
276 |
with gr.Blocks() as demo:
|
277 |
|
278 |
# Top section, greeting and instructions
|
279 |
-
gr.Markdown("##
|
280 |
-
gr.
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
288 |
session_args = gr.State(value=args)
|
289 |
|
290 |
-
with gr.Tab("
|
291 |
|
292 |
with gr.Row():
|
293 |
-
prompt = gr.Textbox(label=f"Prompt", interactive=True)
|
294 |
with gr.Row():
|
295 |
generate_btn = gr.Button("Generate")
|
296 |
with gr.Row():
|
297 |
with gr.Column(scale=2):
|
298 |
-
output_without_watermark = gr.Textbox(label="Output Without Watermark", interactive=False)
|
299 |
with gr.Column(scale=1):
|
300 |
-
without_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False)
|
|
|
301 |
with gr.Row():
|
302 |
with gr.Column(scale=2):
|
303 |
-
output_with_watermark = gr.Textbox(label="Output With Watermark", interactive=False)
|
304 |
with gr.Column(scale=1):
|
305 |
-
with_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False)
|
|
|
306 |
|
307 |
redecoded_input = gr.Textbox(visible=False)
|
308 |
truncation_warning = gr.Number(visible=False)
|
@@ -311,24 +356,16 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
311 |
return redecoded_input + f"\n\n[Prompt was truncated before generation due to length...]", args
|
312 |
else:
|
313 |
return orig_prompt, args
|
314 |
-
|
315 |
-
generate_btn.click(fn=generate_partial, inputs=[prompt,session_args], outputs=[redecoded_input, truncation_warning, output_without_watermark, output_with_watermark,session_args])
|
316 |
-
|
317 |
-
# Show truncated version of prompt if truncation occurred
|
318 |
-
redecoded_input.change(fn=truncate_prompt, inputs=[redecoded_input,truncation_warning,prompt,session_args], outputs=[prompt,session_args])
|
319 |
-
|
320 |
-
# Call detection when the outputs of the generate function are updated.
|
321 |
-
output_without_watermark.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
322 |
-
output_with_watermark.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
323 |
|
324 |
with gr.Tab("Detector Only"):
|
325 |
with gr.Row():
|
326 |
-
|
327 |
-
|
328 |
-
|
|
|
|
|
329 |
with gr.Row():
|
330 |
-
|
331 |
-
detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result, session_args])
|
332 |
|
333 |
# Parameter selection group
|
334 |
with gr.Accordion("Advanced Settings",open=False):
|
@@ -347,18 +384,23 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
347 |
max_new_tokens = gr.Slider(label="Max Generated Tokens", minimum=10, maximum=1000, step=10, value=args.max_new_tokens)
|
348 |
|
349 |
with gr.Column(scale=1):
|
350 |
-
gr.Markdown(f"####
|
351 |
with gr.Row():
|
352 |
gamma = gr.Slider(label="gamma",minimum=0.1, maximum=0.9, step=0.05, value=args.gamma)
|
353 |
with gr.Row():
|
354 |
delta = gr.Slider(label="delta",minimum=0.0, maximum=10.0, step=0.1, value=args.delta)
|
|
|
|
|
|
|
355 |
with gr.Row():
|
356 |
ignore_repeated_bigrams = gr.Checkbox(label="Ignore Bigram Repeats")
|
357 |
with gr.Row():
|
358 |
normalizers = gr.CheckboxGroup(label="Normalizations", choices=["unicode", "homoglyphs", "truecase"], value=args.normalizers)
|
359 |
-
gr.
|
360 |
-
with gr.
|
361 |
-
|
|
|
|
|
362 |
with gr.Accordion("Legacy Settings",open=False):
|
363 |
with gr.Row():
|
364 |
with gr.Column(scale=1):
|
@@ -366,23 +408,31 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
366 |
with gr.Column(scale=1):
|
367 |
select_green_tokens = gr.Checkbox(label="Select 'greenlist' from partition", value=args.select_green_tokens)
|
368 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
369 |
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
|
376 |
-
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
#
|
382 |
def update_sampling_temp(session_state, value): session_state.sampling_temp = float(value); return session_state
|
383 |
def update_generation_seed(session_state, value): session_state.generation_seed = int(value); return session_state
|
384 |
def update_gamma(session_state, value): session_state.gamma = float(value); return session_state
|
385 |
def update_delta(session_state, value): session_state.delta = float(value); return session_state
|
|
|
386 |
def update_decoding(session_state, value):
|
387 |
if value == "multinomial":
|
388 |
session_state.use_sampling = True
|
@@ -405,11 +455,11 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
405 |
def update_normalizers(session_state, value): session_state.normalizers = value; return session_state
|
406 |
def update_seed_separately(session_state, value): session_state.seed_separately = value; return session_state
|
407 |
def update_select_green_tokens(session_state, value): session_state.select_green_tokens = value; return session_state
|
408 |
-
|
409 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[sampling_temp])
|
410 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[generation_seed])
|
411 |
decoding.change(toggle_sampling_vis_inv,inputs=[decoding], outputs=[n_beams])
|
412 |
-
|
413 |
decoding.change(update_decoding,inputs=[session_args, decoding], outputs=[session_args])
|
414 |
sampling_temp.change(update_sampling_temp,inputs=[session_args, sampling_temp], outputs=[session_args])
|
415 |
generation_seed.change(update_generation_seed,inputs=[session_args, generation_seed], outputs=[session_args])
|
@@ -417,17 +467,36 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
417 |
max_new_tokens.change(update_max_new_tokens,inputs=[session_args, max_new_tokens], outputs=[session_args])
|
418 |
gamma.change(update_gamma,inputs=[session_args, gamma], outputs=[session_args])
|
419 |
delta.change(update_delta,inputs=[session_args, delta], outputs=[session_args])
|
|
|
420 |
ignore_repeated_bigrams.change(update_ignore_repeated_bigrams,inputs=[session_args, ignore_repeated_bigrams], outputs=[session_args])
|
421 |
normalizers.change(update_normalizers,inputs=[session_args, normalizers], outputs=[session_args])
|
422 |
seed_separately.change(update_seed_separately,inputs=[session_args, seed_separately], outputs=[session_args])
|
423 |
select_green_tokens.change(update_select_green_tokens,inputs=[session_args, select_green_tokens], outputs=[session_args])
|
424 |
-
|
425 |
generate_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
426 |
detect_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
427 |
|
428 |
-
# When the parameters change, also fire detection, since some detection params dont change the model output.
|
429 |
-
current_parameters.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
430 |
-
current_parameters.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
431 |
|
432 |
demo.queue(concurrency_count=3)
|
433 |
|
|
|
250 |
args)
|
251 |
# decoded_output_with_watermark)
|
252 |
|
253 |
+
def format_names(s):
|
254 |
+
s=s.replace("num_tokens_scored","Tokens Counted (T)")
|
255 |
+
s=s.replace("num_green_tokens","# Tokens in Greenlist")
|
256 |
+
s=s.replace("green_fraction","Fraction of T in Greenlist")
|
257 |
+
s=s.replace("z_score","z-score")
|
258 |
+
s=s.replace("p_value","p value")
|
259 |
+
return s
|
260 |
+
# def str_format_scores(score_dict, detection_threshold):
|
261 |
+
# output_str = f"@ z-score threshold={detection_threshold}:\n\n"
|
262 |
+
# for k,v in score_dict.items():
|
263 |
+
# if k=='green_fraction':
|
264 |
+
# output_str+=f"{format_names(k)}={v:.1%}"
|
265 |
+
# elif k=='confidence':
|
266 |
+
# output_str+=f"{format_names(k)}={v:.3%}"
|
267 |
+
# elif isinstance(v, float):
|
268 |
+
# output_str+=f"{format_names(k)}={v:.3g}"
|
269 |
+
# else:
|
270 |
+
# output_str += v
|
271 |
+
# return output_str
|
272 |
+
def list_format_scores(score_dict, detection_threshold):
|
273 |
+
lst_2d = []
|
274 |
+
lst_2d.append(["z-score threshold", f"{detection_threshold}"])
|
275 |
+
for k,v in score_dict.items():
|
276 |
+
if k=='green_fraction':
|
277 |
+
lst_2d.append([format_names(k), f"{v:.1%}"])
|
278 |
+
elif k=='confidence':
|
279 |
+
lst_2d.append([format_names(k), f"{v:.3%}"])
|
280 |
+
elif isinstance(v, float):
|
281 |
+
lst_2d.append([format_names(k), f"{v:.3g}"])
|
282 |
+
elif isinstance(v, bool):
|
283 |
+
lst_2d.append([format_names(k), ("Watermarked" if v else "Human/Unwatermarked")])
|
284 |
+
else:
|
285 |
+
lst_2d.append([format_names(k), f"{v}"])
|
286 |
+
return lst_2d
|
287 |
+
|
288 |
def detect(input_text, args, device=None, tokenizer=None):
|
289 |
watermark_detector = WatermarkDetector(vocab=list(tokenizer.get_vocab().values()),
|
290 |
gamma=args.gamma,
|
|
|
297 |
select_green_tokens=args.select_green_tokens)
|
298 |
if len(input_text)-1 > watermark_detector.min_prefix_len:
|
299 |
score_dict = watermark_detector.detect(input_text)
|
300 |
+
# output = str_format_scores(score_dict, watermark_detector.z_threshold)
|
301 |
+
output = list_format_scores(score_dict, watermark_detector.z_threshold)
|
302 |
else:
|
303 |
+
# output = (f"Error: string not long enough to compute watermark presence.")
|
304 |
+
output = [["Error","string too short to compute metrics"]]
|
305 |
+
output += [["",""] for _ in range(6)]
|
306 |
+
return output, args
|
307 |
|
308 |
def run_gradio(args, model=None, device=None, tokenizer=None):
|
309 |
|
|
|
313 |
with gr.Blocks() as demo:
|
314 |
|
315 |
# Top section, greeting and instructions
|
316 |
+
gr.Markdown("## 💧 [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226) 🔍")
|
317 |
+
gr.Markdown("[jwkirchenbauer/lm-watermarking![](https://badgen.net/badge/icon/GitHub?icon=github&label)](https://github.com/jwkirchenbauer/lm-watermarking)")
|
318 |
+
|
319 |
+
with gr.Accordion("A note on model capability",open=False):
|
320 |
+
gr.Markdown(
|
321 |
+
"""
|
322 |
+
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.
|
323 |
+
|
324 |
+
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.
|
325 |
+
|
326 |
+
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.
|
327 |
+
"""
|
328 |
+
)
|
329 |
+
|
330 |
+
# Construct state for parameters, define updates and toggles
|
331 |
session_args = gr.State(value=args)
|
332 |
|
333 |
+
with gr.Tab("Generate and Detect"):
|
334 |
|
335 |
with gr.Row():
|
336 |
+
prompt = gr.Textbox(label=f"Prompt", interactive=True,lines=12,max_lines=12)
|
337 |
with gr.Row():
|
338 |
generate_btn = gr.Button("Generate")
|
339 |
with gr.Row():
|
340 |
with gr.Column(scale=2):
|
341 |
+
output_without_watermark = gr.Textbox(label="Output Without Watermark", interactive=False,lines=12,max_lines=12)
|
342 |
with gr.Column(scale=1):
|
343 |
+
# without_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=12,max_lines=12)
|
344 |
+
without_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
345 |
with gr.Row():
|
346 |
with gr.Column(scale=2):
|
347 |
+
output_with_watermark = gr.Textbox(label="Output With Watermark", interactive=False,lines=12,max_lines=12)
|
348 |
with gr.Column(scale=1):
|
349 |
+
# with_watermark_detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=12,max_lines=12)
|
350 |
+
with_watermark_detection_result = gr.Dataframe(headers=["Metric", "Value"],interactive=False,row_count=7,col_count=2)
|
351 |
|
352 |
redecoded_input = gr.Textbox(visible=False)
|
353 |
truncation_warning = gr.Number(visible=False)
|
|
|
356 |
return redecoded_input + f"\n\n[Prompt was truncated before generation due to length...]", args
|
357 |
else:
|
358 |
return orig_prompt, args
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
359 |
|
360 |
with gr.Tab("Detector Only"):
|
361 |
with gr.Row():
|
362 |
+
with gr.Column(scale=2):
|
363 |
+
detection_input = gr.Textbox(label="Text to Analyze", interactive=True,lines=12,max_lines=12)
|
364 |
+
with gr.Column(scale=1):
|
365 |
+
# detection_result = gr.Textbox(label="Detection Result", interactive=False,lines=12,max_lines=12)
|
366 |
+
detection_result = gr.Dataframe(headers=["Metric", "Value"], interactive=False,row_count=7,col_count=2)
|
367 |
with gr.Row():
|
368 |
+
detect_btn = gr.Button("Detect")
|
|
|
369 |
|
370 |
# Parameter selection group
|
371 |
with gr.Accordion("Advanced Settings",open=False):
|
|
|
384 |
max_new_tokens = gr.Slider(label="Max Generated Tokens", minimum=10, maximum=1000, step=10, value=args.max_new_tokens)
|
385 |
|
386 |
with gr.Column(scale=1):
|
387 |
+
gr.Markdown(f"#### Watermark Parameters")
|
388 |
with gr.Row():
|
389 |
gamma = gr.Slider(label="gamma",minimum=0.1, maximum=0.9, step=0.05, value=args.gamma)
|
390 |
with gr.Row():
|
391 |
delta = gr.Slider(label="delta",minimum=0.0, maximum=10.0, step=0.1, value=args.delta)
|
392 |
+
gr.Markdown(f"#### Detector Parameters")
|
393 |
+
with gr.Row():
|
394 |
+
detection_z_threshold = gr.Slider(label="z-score threshold",minimum=0.0, maximum=10.0, step=0.1, value=args.detection_z_threshold)
|
395 |
with gr.Row():
|
396 |
ignore_repeated_bigrams = gr.Checkbox(label="Ignore Bigram Repeats")
|
397 |
with gr.Row():
|
398 |
normalizers = gr.CheckboxGroup(label="Normalizations", choices=["unicode", "homoglyphs", "truecase"], value=args.normalizers)
|
399 |
+
# with gr.Accordion("Actual submitted parameters:",open=False):
|
400 |
+
with gr.Row():
|
401 |
+
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._")
|
402 |
+
with gr.Row():
|
403 |
+
current_parameters = gr.Textbox(label="Current Parameters", value=args)
|
404 |
with gr.Accordion("Legacy Settings",open=False):
|
405 |
with gr.Row():
|
406 |
with gr.Column(scale=1):
|
|
|
408 |
with gr.Column(scale=1):
|
409 |
select_green_tokens = gr.Checkbox(label="Select 'greenlist' from partition", value=args.select_green_tokens)
|
410 |
|
411 |
+
gr.HTML("""
|
412 |
+
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
|
413 |
+
<br/>
|
414 |
+
<a href="https://huggingface.co/spaces/tomg-group-umd/lm-watermarking?duplicate=true">
|
415 |
+
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
416 |
+
<p/>
|
417 |
+
""")
|
418 |
|
419 |
+
# Register main generation tab click, outputing generations as well as a the encoded+redecoded+potentially truncated prompt and flag
|
420 |
+
generate_btn.click(fn=generate_partial, inputs=[prompt,session_args], outputs=[redecoded_input, truncation_warning, output_without_watermark, output_with_watermark,session_args])
|
421 |
+
# Show truncated version of prompt if truncation occurred
|
422 |
+
redecoded_input.change(fn=truncate_prompt, inputs=[redecoded_input,truncation_warning,prompt,session_args], outputs=[prompt,session_args])
|
423 |
+
# Call detection when the outputs (of the generate function) are updated
|
424 |
+
output_without_watermark.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
425 |
+
output_with_watermark.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
426 |
+
# Register main detection tab click
|
427 |
+
detect_btn.click(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_result, session_args])
|
428 |
+
|
429 |
+
# State management logic
|
430 |
+
# update callbacks that change the state dict
|
431 |
def update_sampling_temp(session_state, value): session_state.sampling_temp = float(value); return session_state
|
432 |
def update_generation_seed(session_state, value): session_state.generation_seed = int(value); return session_state
|
433 |
def update_gamma(session_state, value): session_state.gamma = float(value); return session_state
|
434 |
def update_delta(session_state, value): session_state.delta = float(value); return session_state
|
435 |
+
def update_detection_z_threshold(session_state, value): session_state.detection_z_threshold = float(value); return session_state
|
436 |
def update_decoding(session_state, value):
|
437 |
if value == "multinomial":
|
438 |
session_state.use_sampling = True
|
|
|
455 |
def update_normalizers(session_state, value): session_state.normalizers = value; return session_state
|
456 |
def update_seed_separately(session_state, value): session_state.seed_separately = value; return session_state
|
457 |
def update_select_green_tokens(session_state, value): session_state.select_green_tokens = value; return session_state
|
458 |
+
# registering callbacks for toggling the visibilty of certain parameters
|
459 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[sampling_temp])
|
460 |
decoding.change(toggle_sampling_vis,inputs=[decoding], outputs=[generation_seed])
|
461 |
decoding.change(toggle_sampling_vis_inv,inputs=[decoding], outputs=[n_beams])
|
462 |
+
# registering all state update callbacks
|
463 |
decoding.change(update_decoding,inputs=[session_args, decoding], outputs=[session_args])
|
464 |
sampling_temp.change(update_sampling_temp,inputs=[session_args, sampling_temp], outputs=[session_args])
|
465 |
generation_seed.change(update_generation_seed,inputs=[session_args, generation_seed], outputs=[session_args])
|
|
|
467 |
max_new_tokens.change(update_max_new_tokens,inputs=[session_args, max_new_tokens], outputs=[session_args])
|
468 |
gamma.change(update_gamma,inputs=[session_args, gamma], outputs=[session_args])
|
469 |
delta.change(update_delta,inputs=[session_args, delta], outputs=[session_args])
|
470 |
+
detection_z_threshold.change(update_detection_z_threshold,inputs=[session_args, detection_z_threshold], outputs=[session_args])
|
471 |
ignore_repeated_bigrams.change(update_ignore_repeated_bigrams,inputs=[session_args, ignore_repeated_bigrams], outputs=[session_args])
|
472 |
normalizers.change(update_normalizers,inputs=[session_args, normalizers], outputs=[session_args])
|
473 |
seed_separately.change(update_seed_separately,inputs=[session_args, seed_separately], outputs=[session_args])
|
474 |
select_green_tokens.change(update_select_green_tokens,inputs=[session_args, select_green_tokens], outputs=[session_args])
|
475 |
+
# register additional callback on button clicks that updates the shown parameters window
|
476 |
generate_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
477 |
detect_btn.click(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
478 |
+
# When the parameters change, display the update and fire detection, since some detection params dont change the model output.
|
479 |
+
gamma.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
480 |
+
gamma.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
481 |
+
gamma.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
482 |
+
gamma.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_input,session_args])
|
483 |
+
detection_z_threshold.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
484 |
+
detection_z_threshold.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
485 |
+
detection_z_threshold.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
486 |
+
detection_z_threshold.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_input,session_args])
|
487 |
+
ignore_repeated_bigrams.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
488 |
+
ignore_repeated_bigrams.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
489 |
+
ignore_repeated_bigrams.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
490 |
+
ignore_repeated_bigrams.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_input,session_args])
|
491 |
+
normalizers.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
492 |
+
normalizers.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
493 |
+
normalizers.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
494 |
+
normalizers.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_input,session_args])
|
495 |
+
select_green_tokens.change(lambda value: str(value), inputs=[session_args], outputs=[current_parameters])
|
496 |
+
select_green_tokens.change(fn=detect_partial, inputs=[output_without_watermark,session_args], outputs=[without_watermark_detection_result,session_args])
|
497 |
+
select_green_tokens.change(fn=detect_partial, inputs=[output_with_watermark,session_args], outputs=[with_watermark_detection_result,session_args])
|
498 |
+
select_green_tokens.change(fn=detect_partial, inputs=[detection_input,session_args], outputs=[detection_input,session_args])
|
499 |
|
|
|
|
|
|
|
500 |
|
501 |
demo.queue(concurrency_count=3)
|
502 |
|
homoglyph_data/__init__.py
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# This is data for homoglyph finding
|
2 |
+
|
3 |
+
"""Original package info:
|
4 |
+
|
5 |
+
Homoglyphs
|
6 |
+
* Get similar letters
|
7 |
+
* Convert string to ASCII letters
|
8 |
+
* Detect possible letter languages
|
9 |
+
* Detect letter UTF-8 group.
|
10 |
+
|
11 |
+
# main package info
|
12 |
+
__title__ = 'Homoglyphs'
|
13 |
+
__version__ = '2.0.4'
|
14 |
+
__author__ = 'Gram Orsinium'
|
15 |
+
__license__ = 'MIT'
|
16 |
+
|
17 |
+
# License:
|
18 |
+
|
19 |
+
MIT License 2019 orsinium <[email protected]>
|
20 |
+
|
21 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
22 |
+
of this software and associated documentation files (the "Software"), to deal
|
23 |
+
in the Software without restriction, including without limitation the rights
|
24 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
25 |
+
copies of the Software, and to permit persons to whom the Software is
|
26 |
+
furnished to do so, subject to the following conditions:
|
27 |
+
|
28 |
+
The above copyright notice and this permission notice (including the next
|
29 |
+
paragraph) shall be included in all copies or substantial portions of the
|
30 |
+
Software.
|
31 |
+
|
32 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
33 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
34 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
35 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
36 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
37 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
38 |
+
SOFTWARE.
|
39 |
+
|
40 |
+
"""
|
homoglyph_data/categories.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
homoglyph_data/confusables_sept2022.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
homoglyph_data/languages.json
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"ar": "ءآأؤإئابةتثجحخدذرزسشصضطظعغػؼؽؾؿـفقكلمنهوىيًٌٍَُِّ",
|
3 |
+
"be": "ʼЁІЎАБВГДЕЖЗЙКЛМНОПРСТУФХЦЧШЫЬЭЮЯабвгдежзйклмнопрстуфхцчшыьэюяёіў",
|
4 |
+
"bg": "АБВГДЕЖЗИЙКЛМНОПРСТУФХЦЧШЩЪЬЮЯабвгдежзийклмнопрстуфхцчшщъьюя",
|
5 |
+
"ca": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÀÈÉÍÏÒÓÚÜÇàèéíïòóúüç·",
|
6 |
+
"cz": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÁÉÍÓÚÝáéíóúýČčĎďĚěŇňŘřŠšŤťŮůŽž",
|
7 |
+
"da": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÅÆØåæø",
|
8 |
+
"de": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÄÖÜßäöü",
|
9 |
+
"el": "ΪΫΆΈΉΊΌΎΏΑΒΓΔΕΖΗΘΙΚΛΜΝΞΟΠΡΣΤΥΦΧΨΩΐΰϊϋάέήίαβγδεζηθικλμνξοπρςστυφχψωόύώ",
|
10 |
+
"en": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz",
|
11 |
+
"eo": "ABCDEFGHIJKLMNOPRSTUVZabcdefghijklmnoprstuvzĈĉĜĝĤĥĴĵŜŝŬŭ",
|
12 |
+
"es": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÁÉÍÑÓÚÜáéíñóúü",
|
13 |
+
"et": "ABDEGHIJKLMNOPRSTUVabdeghijklmnoprstuvÄÕÖÜäõöü",
|
14 |
+
"fi": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÄÅÖäåöŠšŽž",
|
15 |
+
"fr": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÀÂÇÈÉÊÎÏÙÛàâçèéêîïùûŒœ",
|
16 |
+
"he": "אבגדהוזחטיךכלםמןנסעףפץצקרשתװױײ",
|
17 |
+
"hr": "ABCDEFGHIJKLMNOPRSTUVZabcdefghijklmnoprstuvzĆćČčĐ𩹮ž",
|
18 |
+
"hu": "ABCDEFGHIJKLMNOPRSTUVZabcdefghijklmnoprstuvzÁÉÍÓÖÚÜáéíóöúüŐőŰű",
|
19 |
+
"it": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÀÈÉÌÒÓÙàèéìòóù",
|
20 |
+
"lt": "ABCDEFGHIJKLMNOPRSTUVYZabcdefghijklmnoprstuvyzĄąČčĖėĘęĮįŠšŪūŲųŽž",
|
21 |
+
"lv": "ABCDEFGHIJKLMNOPRSTUVZabcdefghijklmnoprstuvzĀāČčĒēĢģĪīĶķĻļŅņŠšŪūŽž",
|
22 |
+
"mk": "ЃЅЈЉЊЌЏАБВГДЕЖЗИКЛМНОПРСТУФХЦЧШабвгдежзиклмнопрстуфхцчшѓѕјљњќџ",
|
23 |
+
"nl": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz",
|
24 |
+
"pl": "ABCDEFGHIJKLMNOPRSTUWYZabcdefghijklmnoprstuwyzÓóĄąĆćĘꣳŃńŚśŹźŻż",
|
25 |
+
"pt": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÀÁÂÃÇÉÊÍÓÔÕÚàáâãçéêíóôõú",
|
26 |
+
"ro": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÂÎâîĂăȘșȚț",
|
27 |
+
"ru": "ЁАБВГДЕЖЗИЙКЛМНОПРСТУФХЦЧШЩЪЫЬЭЮЯабвгдежзийклмнопрстуфхцчшщъыьэюяё",
|
28 |
+
"sk": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzÁÄÉÍÓÔÚÝáäéíóôúýČčĎďĹ弾ŇňŔ੹ŤťŽž",
|
29 |
+
"sl": "ABCDEFGHIJKLMNOPRSTUVZabcdefghijklmnoprstuvzČ芚Žž",
|
30 |
+
"sr": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzЂЈЉЊЋЏАБВГДЕЖЗИКЛМНОПРСТУФХЦЧШабвгдежзиклмнопрстуфхцчшђјљњћџ",
|
31 |
+
"th": "กขฃคฅฆงจฉชซฌญฎฏฐฑฒณดตถทธนบปผฝพฟภมยรฤลฦวศษสหฬอฮฯะัาำิีึืฺุู฿เแโใไๅๆ็่้๊๋์ํ๎๏๐๑๒๓๔๕๖๗๘๙๚๛",
|
32 |
+
"tr": "ABCDEFGHIJKLMNOPRSTUVYZabcdefghijklmnoprstuvyzÂÇÎÖÛÜâçîöûüĞğİıŞş",
|
33 |
+
"vi": "ABCDEGHIKLMNOPQRSTUVXYabcdeghiklmnopqrstuvxyÂÊÔâêôĂăĐđƠơƯư"
|
34 |
+
}
|
homoglyphs.py
CHANGED
@@ -9,10 +9,6 @@ from itertools import product
|
|
9 |
import os
|
10 |
import unicodedata
|
11 |
|
12 |
-
import homoglyphs_fork as hg
|
13 |
-
|
14 |
-
CURRENT_DIR = hg.core.CURRENT_DIR
|
15 |
-
|
16 |
# Actions if char not in alphabet
|
17 |
STRATEGY_LOAD = 1 # load category for this char
|
18 |
STRATEGY_IGNORE = 2 # add char to result
|
@@ -21,13 +17,17 @@ STRATEGY_REMOVE = 3 # remove char from result
|
|
21 |
ASCII_RANGE = range(128)
|
22 |
|
23 |
|
|
|
|
|
|
|
|
|
24 |
class Categories:
|
25 |
"""
|
26 |
Work with aliases from ISO 15924.
|
27 |
https://en.wikipedia.org/wiki/ISO_15924#List_of_codes
|
28 |
"""
|
29 |
|
30 |
-
fpath = os.path.join(
|
31 |
|
32 |
@classmethod
|
33 |
def _get_ranges(cls, categories):
|
@@ -70,8 +70,9 @@ class Categories:
|
|
70 |
# try detect category by unicodedata
|
71 |
try:
|
72 |
category = unicodedata.name(char).split()[0]
|
73 |
-
except TypeError:
|
74 |
# In Python2 unicodedata.name raise error for non-unicode chars
|
|
|
75 |
pass
|
76 |
else:
|
77 |
if category in data["aliases"]:
|
@@ -91,7 +92,7 @@ class Categories:
|
|
91 |
|
92 |
|
93 |
class Languages:
|
94 |
-
fpath = os.path.join(
|
95 |
|
96 |
@classmethod
|
97 |
def get_alphabet(cls, languages):
|
@@ -167,8 +168,7 @@ class Homoglyphs:
|
|
167 |
@staticmethod
|
168 |
def get_table(alphabet):
|
169 |
table = defaultdict(set)
|
170 |
-
|
171 |
-
with open(os.path.join("confusables_sept2022.json")) as f:
|
172 |
data = json.load(f)
|
173 |
for char in alphabet:
|
174 |
if char in data:
|
@@ -180,8 +180,7 @@ class Homoglyphs:
|
|
180 |
@staticmethod
|
181 |
def get_restricted_table(source_alphabet, target_alphabet):
|
182 |
table = defaultdict(set)
|
183 |
-
|
184 |
-
with open(os.path.join("confusables_sept2022.json")) as f:
|
185 |
data = json.load(f)
|
186 |
for char in source_alphabet:
|
187 |
if char in data:
|
@@ -244,9 +243,7 @@ class Homoglyphs:
|
|
244 |
alt_chars = self._get_char_variants(char)
|
245 |
|
246 |
if ascii:
|
247 |
-
alt_chars = [
|
248 |
-
char for char in alt_chars if ord(char) in self.ascii_range
|
249 |
-
]
|
250 |
if not alt_chars and self.ascii_strategy == STRATEGY_IGNORE:
|
251 |
return
|
252 |
|
|
|
9 |
import os
|
10 |
import unicodedata
|
11 |
|
|
|
|
|
|
|
|
|
12 |
# Actions if char not in alphabet
|
13 |
STRATEGY_LOAD = 1 # load category for this char
|
14 |
STRATEGY_IGNORE = 2 # add char to result
|
|
|
17 |
ASCII_RANGE = range(128)
|
18 |
|
19 |
|
20 |
+
CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
|
21 |
+
DATA_LOCATION = os.path.join(CURRENT_DIR, "homoglyph_data")
|
22 |
+
|
23 |
+
|
24 |
class Categories:
|
25 |
"""
|
26 |
Work with aliases from ISO 15924.
|
27 |
https://en.wikipedia.org/wiki/ISO_15924#List_of_codes
|
28 |
"""
|
29 |
|
30 |
+
fpath = os.path.join(DATA_LOCATION, "categories.json")
|
31 |
|
32 |
@classmethod
|
33 |
def _get_ranges(cls, categories):
|
|
|
70 |
# try detect category by unicodedata
|
71 |
try:
|
72 |
category = unicodedata.name(char).split()[0]
|
73 |
+
except (TypeError, ValueError):
|
74 |
# In Python2 unicodedata.name raise error for non-unicode chars
|
75 |
+
# Python3 raise ValueError for non-unicode characters
|
76 |
pass
|
77 |
else:
|
78 |
if category in data["aliases"]:
|
|
|
92 |
|
93 |
|
94 |
class Languages:
|
95 |
+
fpath = os.path.join(DATA_LOCATION, "languages.json")
|
96 |
|
97 |
@classmethod
|
98 |
def get_alphabet(cls, languages):
|
|
|
168 |
@staticmethod
|
169 |
def get_table(alphabet):
|
170 |
table = defaultdict(set)
|
171 |
+
with open(os.path.join(DATA_LOCATION, "confusables_sept2022.json")) as f:
|
|
|
172 |
data = json.load(f)
|
173 |
for char in alphabet:
|
174 |
if char in data:
|
|
|
180 |
@staticmethod
|
181 |
def get_restricted_table(source_alphabet, target_alphabet):
|
182 |
table = defaultdict(set)
|
183 |
+
with open(os.path.join(DATA_LOCATION, "confusables_sept2022.json")) as f:
|
|
|
184 |
data = json.load(f)
|
185 |
for char in source_alphabet:
|
186 |
if char in data:
|
|
|
243 |
alt_chars = self._get_char_variants(char)
|
244 |
|
245 |
if ascii:
|
246 |
+
alt_chars = [char for char in alt_chars if ord(char) in self.ascii_range]
|
|
|
|
|
247 |
if not alt_chars and self.ascii_strategy == STRATEGY_IGNORE:
|
248 |
return
|
249 |
|
requirements.txt
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
homoglyphs_fork
|
2 |
nltk
|
3 |
scipy
|
4 |
torch
|
|
|
|
|
1 |
nltk
|
2 |
scipy
|
3 |
torch
|
watermark_processor.py
CHANGED
@@ -216,6 +216,8 @@ class WatermarkDetector(WatermarkBase):
|
|
216 |
score_dict.update(dict(num_tokens_scored=num_tokens_scored))
|
217 |
if return_num_green_tokens:
|
218 |
score_dict.update(dict(num_green_tokens=green_token_count))
|
|
|
|
|
219 |
if return_z_score:
|
220 |
score_dict.update(dict(z_score=self._compute_z_score(green_token_count, num_tokens_scored)))
|
221 |
if return_p_value:
|
@@ -223,8 +225,6 @@ class WatermarkDetector(WatermarkBase):
|
|
223 |
if z_score is None:
|
224 |
z_score = self._compute_z_score(green_token_count, num_tokens_scored)
|
225 |
score_dict.update(dict(p_value=self._compute_p_value(z_score)))
|
226 |
-
if return_green_fraction:
|
227 |
-
score_dict.update(dict(green_fraction=(green_token_count / num_tokens_scored)))
|
228 |
if return_green_token_mask:
|
229 |
score_dict.update(dict(green_token_mask=green_token_mask))
|
230 |
|
|
|
216 |
score_dict.update(dict(num_tokens_scored=num_tokens_scored))
|
217 |
if return_num_green_tokens:
|
218 |
score_dict.update(dict(num_green_tokens=green_token_count))
|
219 |
+
if return_green_fraction:
|
220 |
+
score_dict.update(dict(green_fraction=(green_token_count / num_tokens_scored)))
|
221 |
if return_z_score:
|
222 |
score_dict.update(dict(z_score=self._compute_z_score(green_token_count, num_tokens_scored)))
|
223 |
if return_p_value:
|
|
|
225 |
if z_score is None:
|
226 |
z_score = self._compute_z_score(green_token_count, num_tokens_scored)
|
227 |
score_dict.update(dict(p_value=self._compute_p_value(z_score)))
|
|
|
|
|
228 |
if return_green_token_mask:
|
229 |
score_dict.update(dict(green_token_mask=green_token_mask))
|
230 |
|