Spaces:
Running
Running
File size: 42,073 Bytes
f7acd50 f655011 4da1fb0 2c5c709 f655011 2c5c709 286d976 fb1c1ed f655011 129904f 95b2105 f655011 2b78064 f655011 2b78064 5a498e2 95b2105 2b78064 f655011 2b78064 f655011 85e58bb c50cfb4 85e58bb 2b78064 f7acd50 f655011 fb1c1ed f655011 9acb8e6 2c5c709 f655011 d646867 f655011 2b78064 f655011 f7a1bd4 f655011 2b78064 f655011 2b78064 f7a1bd4 2b78064 f7a1bd4 2b78064 2c5c709 f7a1bd4 85e58bb f7acd50 d646867 2c5c709 f7acd50 2c5c709 f7acd50 129904f f7acd50 2c5c709 3379490 f7a1bd4 f7acd50 f7a1bd4 129904f 2b78064 129904f f655011 129904f 9acb8e6 3379490 9acb8e6 453c7fc 9acb8e6 85e58bb 2fd6955 85e58bb f7a1bd4 85e58bb f7a1bd4 453c7fc f7a1bd4 453c7fc 2b78064 f7a1bd4 453c7fc f655011 f7a1bd4 d646867 85e58bb f655011 2c5c709 d646867 85e58bb d646867 85e58bb 2c5c709 85e58bb d646867 f655011 3379490 f7a1bd4 f7acd50 f7a1bd4 f655011 3379490 2fd6955 3379490 f7a1bd4 f7acd50 3379490 f7a1bd4 3379490 f7a1bd4 3379490 f7a1bd4 3379490 f7a1bd4 f7acd50 3379490 f7a1bd4 3379490 f7a1bd4 3379490 f7acd50 3379490 f7a1bd4 3379490 f7a1bd4 3379490 f7acd50 3379490 f7a1bd4 3379490 f7acd50 3379490 f7a1bd4 3379490 f7a1bd4 f7acd50 3379490 f7acd50 f7a1bd4 3379490 f7a1bd4 3379490 f7acd50 3379490 f7a1bd4 3379490 85e58bb 2c5c709 2b78064 2c5c709 fb1c1ed 2c5c709 2b78064 2c5c709 fb1c1ed 2c5c709 f7acd50 2c5c709 fb1c1ed 2c5c709 3379490 129904f 3379490 129904f 3379490 f7a1bd4 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 129904f f655011 2c5c709 f655011 2c5c709 f655011 2c5c709 9acb8e6 2c5c709 3379490 2c5c709 3379490 2c5c709 129904f 2c5c709 3379490 2c5c709 129904f 2c5c709 f7acd50 2c5c709 d646867 2c5c709 d646867 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 2c5c709 f7acd50 f655011 2c5c709 f7acd50 2c5c709 fb1c1ed 2c5c709 fb1c1ed 2c5c709 f7acd50 3379490 2c5c709 f7acd50 3379490 2c5c709 3379490 2c5c709 3379490 f7acd50 5f7f1fd f655011 f7acd50 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 |
"""
Controlled Chat is a graphical and chat interface to Representation Engineering.
It creates a single Gradio application to be run locally or on a Hugging Face space.
This version is intended to run on CPU, and so uses Llama 3.2 1B.
It is hosted online at https://huggingface.co/spaces/Abrak/Controlled_Chat_CPU/.
There is also a GPU version based on Mistral 0.3 9B, requiring 16GB of VRAM.
Find it at https://huggingface.co/spaces/Abrak/Controlled_Chat.
You can also run thie application locally: create a venv, install the requirements, and run this script.
If you want to port this to another model, you'll need to do a few things:
1. Change the model path on the first line of code
2. Experiment with different ranges of layers in the call to ControlModel()
3. Change out the construct_prompt_* function to fit the model's prompt syntax
4. Call train_models()
If you clone this project, you can add new models into the control_models directory and everyting should work.
This file's code is licensed under MIT. See the README.MD and LLAMA LICENSE.TXT.
"""
import os
import threading
import json
import csv
import torch
import re
import tempfile
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from repeng import ControlVector, ControlModel, DatasetEntry
import gradio as gr
# Initialize model and tokenizer
from huggingface_hub import login
# Initialize model and tokenizer
llama_path = "meta-llama/Llama-3.2-1B-Instruct"
#llama_path = r"E:/language_models/models/mistral"
access_token = os.getenv("llamaaccesstoken")
login(access_token)
tokenizer = AutoTokenizer.from_pretrained(llama_path)
tokenizer.pad_token_id = 0
model = AutoModelForCausalLM.from_pretrained(
llama_path,
torch_dtype=torch.float16,
trust_remote_code=True,
use_safetensors=True
)
cuda = torch.cuda.is_available()
print(f"Is CUDA available: {cuda}")
model = model.to("cuda:0" if cuda else "cpu")
if cuda:
print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")
# in mistral, there are 32 layers from -31 to 0. set to 13 layers from -5 to -18
# model = ControlModel(model, list(range(-5, -18, -1)))
# in llama 3.2 there are 32 layers from 0 to 15. With some experimentation, I found setting layers 10 through 5 is best
model = ControlModel(model, list(range(10, 5, -1)))
# Generation settings
# Generation settings
default_generation_settings = {
"pad_token_id": tokenizer.eos_token_id,
"do_sample": False, # Deterministic output
"max_new_tokens": 384,
"repetition_penalty": 1.1, # Reduce repetition
}
# List available control vectors
control_vector_files = [f for f in os.listdir('control_models') if f.endswith('.gguf')]
if not control_vector_files:
pass
#raise FileNotFoundError("No .gguf control vector files found in the control_models directory.")
# Function to toggle slider visibility based on checkbox state
def toggle_slider(checked):
return gr.update(visible=checked)
def construct_prompt_mistral(history, system_prompt, user_message):
"""
Converts the history (list of tuples) back into the string format Mistral expects
"""
formatted_prompt = ""
user_tag, asst_tag = "[INST]", "[/INST]"
# <s>[INST] user message[/INST] assistant message</s>[INST] new user message[/INST]
# Mistral expects the history to be wrapped in <s>history</s>, so it's added here
if len(history) > 0:
formatted_prompt += "<s>"
# Append the system prompt if provided
if system_prompt.strip():
formatted_prompt += f"{user_tag} {system_prompt}{asst_tag} "
# Construct the formatted prompt based on history
if len(history) > 0:
for turn in history:
user_msg, asst_msg = turn
asst_msg = asst_msg.split("\n")[1:]
formatted_prompt += f"{user_tag} {user_msg} {asst_tag} {asst_msg}"
if len(history) > 0:
formatted_prompt += "</s>"
# Append the new user message
formatted_prompt += f"{user_tag} {user_message} {asst_tag}"
return formatted_prompt
def construct_prompt_llama(history, system_prompt, user_message):
"""
Converts the history (list of tuples) back into the string format LLama expects
LLama prompt format:
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
Cutting Knowledge Date: December 2023
Today Date: 23 July 2024
You are a helpful assistant
<|eot_id|>
<|start_header_id|>user<|end_header_id|>
What is the capital of France?
<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
"""
formatted_prompt = ""
# Begin the prompt with the start token
formatted_prompt += "<|begin_of_text|>\n"
# Append the system prompt if provided
if system_prompt.strip():
formatted_prompt += "<|start_header_id|>system<|end_header_id|>\n"
formatted_prompt += f"{system_prompt.strip()}"
formatted_prompt += "<|eot_id|>\n"
# Construct the formatted prompt based on history
for user_msg, asst_msg in history:
# Append the user message
formatted_prompt += "<|start_header_id|>user<|end_header_id|>\n"
formatted_prompt += f"{user_msg.strip()}"
formatted_prompt += "<|eot_id|>\n"
# Append the assistant's response
formatted_prompt += "<|start_header_id|>assistant<|end_header_id|>\n"
formatted_prompt += f"{asst_msg.strip()}"
formatted_prompt += "<|eot_id|>\n"
# Append the new user message
formatted_prompt += "<|start_header_id|>user<|end_header_id|>\n"
formatted_prompt += f"{user_message.strip()}"
formatted_prompt += "<|eot_id|>\n"
# Indicate that the assistant should provide a response
formatted_prompt += "<|start_header_id|>assistant<|end_header_id|>\n"
return formatted_prompt
def generate_response(system_prompt, user_message, history, max_new_tokens, repitition_penalty, do_sample, user_model, input_checkbox, input_slider, *args):
"""
Applies the control vectors and calls the language model.
Returns a list of tuples, the user message and the assistant response,
which Gradio uses to update the chatbot history
"""
global previous_turn
previous_turn = user_message
combined_vector = None
assistant_message_title = ""
# args not included in test_generate
if args:
# Separate checkboxes and sliders based on type
# The first x in args are the checkbox names (the file names)
# The second x in args are the slider values
checkboxes = []
sliders = []
for i in range(len(control_vector_files)):
checkboxes.append(args[i])
sliders.append(args[len(control_vector_files) + i])
# Apply selected control vectors with their corresponding weights
control_vectors = []
for i in range(len(control_vector_files)):
if checkboxes[i]:
cv_file = control_vector_files[i]
weight = sliders[i]
# Set the control vector's weight (and sign) by multiplying by its slider value
control_vectors.append(ControlVector.import_gguf(f"control_models/{cv_file}") * weight)
assistant_message_title += f"{cv_file.split('.')[0]}: {weight};"
# The control model takes a sum of positive and negative control vectors
for i in range(len(control_vectors)):
if combined_vector is None:
combined_vector = control_vectors[i]
else:
combined_vector += control_vectors[i]
if input_checkbox:
# User has uploaded their own gguf control vector
input_vector = ControlVector.import_gguf(user_model)
if combined_vector is None:
combined_vector = input_vector * input_slider
else:
combined_vector += input_vector * input_slider
assistant_message_title += f"Uploaded: {input_slider};"
# Set the combined set of vectors as the control for the model
try:
if combined_vector is not None:
model.reset()
model.set_control(combined_vector)
except Exception as e:
print(f"Failed to set Control: {e}")
formatted_prompt = construct_prompt_llama(history, system_prompt, user_message)
# Tokenize the input
input_ids = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
generation_settings = {
"pad_token_id": tokenizer.eos_token_id,
"do_sample": do_sample,
"max_new_tokens": int(max_new_tokens),
"repetition_penalty": repetition_penalty.value,
}
timeout = 120.0
if cuda:
timeout = 15.0
_streamer = TextIteratorStreamer(tokenizer, timeout=timeout, skip_prompt=True, skip_special_tokens=False,)
generate_kwargs = dict(
input_ids,
streamer=_streamer,
pad_token_id= tokenizer.eos_token_id,
do_sample= do_sample,
max_new_tokens= int(max_new_tokens),
repetition_penalty= repetition_penalty.value,
)
t = threading.Thread(target=model.generate, kwargs=generate_kwargs)
t.start()
# Display the response as it streams in, prepending the control vector info
partial_message = ""
#show the control vector info while we wait for the first token
temp_output = "*" + assistant_message_title + "*" + "\n\n*Please wait*..." + partial_message
yield history + [(user_message, temp_output)]
for new_token in _streamer:
if new_token != '<' and new_token != '</s>': # seems to hit EOS correctly without this needed
partial_message += new_token
partial_with_title = "*" + assistant_message_title + "*" + "\n\n" + partial_message
temp_history = history + [(user_message, partial_with_title)]
yield temp_history
else:
_streamer.end()
# remove the trailing </s> if present
# it won't be present if the model ran out from max_tokens
def get_assistant_response(input_string):
if len(input_string) >= 10:
if input_string[-10:] == "<|eot_id|>":
return input_string[:-10]
else:
return input_string
else:
return input_string
# Update conversation history
assistant_response = get_assistant_response(partial_message)
assistant_response_display = f"*{assistant_message_title}*\n\n{assistant_response}"
# Update conversation history
history.append((user_message, assistant_response_display))
return history
def generate_response_with_retry(system_prompt, user_message, history, max_new_tokens, repitition_penalty, do_sample, user_model, input_checkbox, input_slider, *args):
# Remove last user input and assistant response from history, then call generate_response()
global previous_turn
previous_ueser_message = previous_turn
if history:
history = history[0:-1]
# Using the previous turn's text, even though it isn't in the textbox anymore
for output in generate_response(system_prompt, previous_ueser_message, history, max_new_tokens, repetition_penalty, do_sample, user_model, input_checkbox, input_slider, *args):
yield [output, previous_ueser_message]
# Function to reset the conversation history
def reset_chat():
# returns a blank state
return [], ""
def get_checkboxes():
# rebuilding the list of checkboxes, so that these presets don't have to change
# when adding a new control model
# Warning: adding any new components into the header before the checkboxes is going to break this path
checkbox_column = app.children[0].children[0].children[2].children[0].children
#checkbox_column = app.children[2].children[0].children
model_names_and_indexes = {}
checkbox_index = 0
for i in range(len(checkbox_column)):
if isinstance(checkbox_column[i], gr.Row):
try:
model_name = checkbox_column[i].children[0].children[0].label
model_names_and_indexes[model_name] = checkbox_index
checkbox_index += 1
except IndexError:
# allow for other rows to be in the interface
pass
except AttributeError:
pass
return model_names_and_indexes
def set_preset_helpful(*args):
# gets the list of all checkboxes and sliders
# sets checkboxes and sliders accordingly to this persona
# args is a list of checkboxes and then slider values
# must return the updated list of checkboxes and sliders
new_checkbox_values = []
new_slider_values = []
model_names_and_indexes = get_checkboxes()
for check in model_names_and_indexes:
if check == "Empathetic":
new_checkbox_values.append(True)
new_slider_values.append(1.0)
elif check == "Optimistic":
new_checkbox_values.append(True)
new_slider_values.append(1.0)
else:
new_checkbox_values.append(False)
new_slider_values.append(0.0)
return new_checkbox_values + new_slider_values
def set_preset_conspiracist(*args):
# gets the list of all checkboxes and sliders
# sets checkboxes and sliders accordingly to this persona
# args is a list of checkboxes and then slider values
# must return the updated list of checkboxes and sliders
new_checkbox_values = []
new_slider_values = []
model_names_and_indexes = get_checkboxes()
for check in model_names_and_indexes:
if check == "Conspiracist":
new_checkbox_values.append(True)
new_slider_values.append(1.5)
elif check == "Creative":
new_checkbox_values.append(True)
new_slider_values.append(1.0)
elif check == "Lazy":
new_checkbox_values.append(True)
new_slider_values.append(-0.5)
elif check == "Honest":
new_checkbox_values.append(True)
new_slider_values.append(-1.0)
else:
new_checkbox_values.append(False)
new_slider_values.append(0.0)
return new_checkbox_values + new_slider_values
def set_preset_stoner(*args):
# gets the list of all checkboxes and sliders
# sets checkboxes and sliders accordingly to this persona
# args is a list of checkboxes and then slider values
# must return the updated list of checkboxes and sliders
new_checkbox_values = []
new_slider_values = []
model_names_and_indexes = get_checkboxes()
for check in model_names_and_indexes:
if check == "Angry":
new_checkbox_values.append(True)
new_slider_values.append(0.3)
elif check == "Conservative":
new_checkbox_values.append(True)
new_slider_values.append(-0.5)
elif check == "Tripping":
new_checkbox_values.append(True)
new_slider_values.append(1.0)
else:
new_checkbox_values.append(False)
new_slider_values.append(0.0)
return new_checkbox_values + new_slider_values
def set_preset_facts(*args):
# gets the list of all checkboxes and sliders
# sets checkboxes and sliders accordingly to this persona
# args is a list of checkboxes and then slider values
# must return the updated list of checkboxes and sliders
new_checkbox_values = []
new_slider_values = []
model_names_and_indexes = get_checkboxes()
for check in model_names_and_indexes:
if check == "Worried":
new_checkbox_values.append(True)
new_slider_values.append(-0.5)
elif check == "Joking":
new_checkbox_values.append(True)
new_slider_values.append(-0.5)
elif check == "Lazy":
new_checkbox_values.append(True)
new_slider_values.append(-0.5)
elif check == "Honest":
new_checkbox_values.append(True)
new_slider_values.append(0.5)
else:
new_checkbox_values.append(False)
new_slider_values.append(0.0)
return new_checkbox_values + new_slider_values
def disable_controls():
return gr.update(interactive= False, value= "⌛ Processing"), gr.update(interactive=False)
def enable_controls():
return gr.update(interactive= True, value= "💬 Submit"), gr.update(interactive= True)
def clear_input(input_textbox):
return ""
def make_dataset(
template: str,
positive_personas: list[str],
negative_personas: list[str],
suffix_list: list[str]
) -> list[DatasetEntry]:
dataset = []
# Tags for prompt formatting with Llama
user_tag = "<|start_header_id|>user<|end_header_id|>\n\n"
asst_tag = "<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>"
for suffix in suffix_list:
for positive_persona, negative_persona in zip(positive_personas, negative_personas):
positive_template = template.format(persona=positive_persona)
negative_template = template.format(persona=negative_persona)
dataset.append(
DatasetEntry(
positive=f"{user_tag} {positive_template} {asst_tag} {suffix}",
negative=f"{user_tag} {negative_template} {asst_tag} {suffix}",
)
)
return dataset
def train_model_persona(positive_text, negative_text):
positive_list = positive_text.split('\n')
negative_list = negative_text.split('\n')
with open("all_truncated_outputs.json") as f:
output_suffixes = json.load(f)
dataset = make_dataset(
"Act as if you are an extremely {persona} person",
positive_list,
negative_list,
output_suffixes)
model.reset()
output_model = ControlVector.train(model, tokenizer, dataset)
# Write file to temporary directory returning the path to Gradio for download
filename = re.sub(r'[ <>:"/\\|?*]', '', positive_list[0]) + '_'
temp_file = tempfile.NamedTemporaryFile(
prefix=filename,
suffix=".gguf",
delete= False
)
ControlVector.export_gguf(output_model, temp_file.name)
temp_file.close()
return temp_file.name
def train_model_facts(positive_text, negative_text):
with open("true_facts.csv") as f:
reader = csv.reader(f)
fact_suffixes = list(reader)
dataset = make_dataset(
"Pretend to be a {persona} making statements about the world.",
positive_text,
negative_text,
fact_suffixes
)
output_model = ControlVector.train(model, tokenizer, dataset)
filename = re.sub(r'[ <>:"/\\|?*]', '', positive_text) + '_'
temp_file = tempfile.NamedTemporaryFile(
prefix=filename,
suffix=".gguf",
delete= False
)
ControlVector.export_gguf(output_model, temp_file.name)
temp_file.close()
return temp_file.name
tooltip_css = """
/* Tooltip container */
.tooltip {
position: relative;
display: inline-block;
cursor: help;
}
/* Tooltip text */
.tooltip .tooltiptext {
visibility: hidden;
width: 200px;
background-color: #1f2937;
color: #f3f4f6;
text-align: left;
border-radius: 6px;
padding: 8px;
position: absolute;
z-index: 1;
bottom: 125%; /* Position above the element */
left: 50%;
margin-left: -100px;
opacity: 0;
transition: opacity 0.3s;
}
/* Tooltip arrow */
.tooltip .tooltiptext::after {
content: "";
position: absolute;
top: 100%; /* At the bottom of tooltip */
left: 50%;
margin-left: -5px;
border-width: 5px;
border-style: solid;
border-color: #1f2937 transparent transparent transparent;
}
/* Show the tooltip text when hovering */
.tooltip:hover .tooltiptext {
visibility: visible;
opacity: 1;"""
dark_theme = gr.Theme.from_hub("ParityError/Anime").set(
# body_background_fill= "url(https://image uri) #000000 no-repeat right bottom / auto 100svh padding-box fixed;",
# body_background_fill_dark= "url(https://image uri) #000000 no-repeat right bottom / auto 100svh padding-box fixed;",
)
with gr.Blocks(
theme=dark_theme,
css=tooltip_css,
) as app:
with gr.Tab(
label="Use"
):
# Header
if cuda:
gr.Markdown("# 🧠 LLM Mind Control (Llama 3.2 1B)")
else:
gr.Markdown("""# 🧠 LLM Mind Control ((Llama 3.2 1B))
*Warning: although using a small model, running on CPU will still be very slow*""")
gr.Markdown("""Unlike prompting, direct weight manipulation lets you fine-tune the amount of a personality
trait or topic. Enabled through [Representation Engineering](https://arxiv.org/abs/2310.01405)
via the [repeng](https://pypi.org/project/repeng) library.
[Watch a demo](https://youtu.be/gYZPGVafD7M) for usage tips.""")
with gr.Row():
# Left Column: Control Vectors and advanced settings
with gr.Column(scale=1):
gr.Markdown("### ⚡ Control Vectors")
control_vector_label = gr.HTML("""
<div class="tooltip">
<span>Select how you want to control the LLM per turn - towards (+) or away (-). Or start with a preset:</span>
<span class="tooltiptext">+/- 1.0 is a good start. Check the examples for each vector.</span>
</div>
""")
with gr.Row():
button_helpful = gr.Button(
value="Kind and helpful",
)
button_facts = gr.Button(
value="Just the facts"
)
button_stoner = gr.Button(
value="Angry stoner"
)
button_conspiracist = gr.Button(
value="Manic conspiracist"
)
# Create checkboxes and sliders for each control vector
control_checks = []
control_sliders = []
for cv_file in control_vector_files:
with gr.Row():
# Checkbox to select the control vector
checkbox = gr.Checkbox(label=cv_file.split('.')[0], value=False)
control_checks.append(checkbox)
# Slider to adjust the control vector's weight
slider = gr.Slider(
minimum=-2.5,
maximum=2.5,
value=0.0,
step=0.1,
label=f"Voltage",
visible=False
)
control_sliders.append(slider)
# Link the checkbox to toggle slider visibility
checkbox.change(
toggle_slider,
inputs=checkbox,
outputs=slider
)
# Upload your own control model
with gr.Accordion("📎 Use your own model", open=False):
with gr.Row():
input_model = gr.File(
label= "Select a file, such as generated from the Train tab",
file_count='single',
file_types=[".gguf"]
)
input_model_checkbox = gr.Checkbox(
value= False,
label= "Use uploaded model"
)
input_model_slider = gr.Slider(
minimum=-2.5,
maximum=2.5,
value=0.0,
step=0.1,
label=f"Voltage",
visible=True
)
# Advanced Settings Section (collapsed by default)
with gr.Accordion("🔧 Advanced Settings", open=False):
with gr.Row():
system_prompt = gr.Textbox(
lines=2,
value="Respond to the user concisely",
interactive=True,
label="System Prompt",
show_label=False
)
# Max Response Length with tooltip
with gr.Column(scale=1):
max_tokens_label = gr.HTML("""
<div class="tooltip">
<span>Max Response Length (in tokens)</span>
<span class="tooltiptext">Lower for faster output, higher to allow longer answers</span>
</div>
""")
max_new_tokens = gr.Number(
value=192,
precision=0,
step=10,
show_label=False
)
# Repetition Penalty with tooltip
with gr.Column(scale=1):
repetition_label = gr.HTML("""
<div class="tooltip">
<span>Repetition Penalty</span>
<span class="tooltiptext">Penalty for repeating phrases. Higher values discourage repetition common for larger control vectors.</span>
</div>
""")
repetition_penalty = gr.Number(
value=1.1,
precision=2,
step=0.1,
show_label=False
)
# Non-deterministic output with tooltip
with gr.Column(scale=1):
do_sample_label = gr.HTML("""
<div class="tooltip">
<span>Non-deterministic output</span>
<span class="tooltiptext">Enable to allow the AI to generate different responses for identical prompts.</span>
</div>
""")
do_sample = gr.Checkbox(
value=False,
show_label=False,
label="do_sample"
)
toggle_dark = gr.Button(value="Toggle Dark Mode")
gr.Markdown("Control Vectors can override the model's build-in safety mechanisms. Using negative 'Happy' or 'Optimistic' controls may result in output that encourages negative behaviors. Use at your own risk.")
gr.Markdown("Built with Llama. See LLAMA LICENSE.txt")
# Right Column: Chat Interface
with gr.Column(scale=2):
gr.Markdown("### 🗨️ Conversation")
# Chatbot to display conversation
chatbot = gr.Chatbot(
type="tuples"
)
# User Message Input with tooltip
#with gr.Row():
user_input_label = gr.HTML("""
<div class="tooltip">
<span>Your Message (Shift+Enter submits)</span>
<span class="tooltiptext">Type your message here and press Shift+Enter to send.</span>
</div>
""")
user_input = gr.Textbox(
lines=2,
placeholder="I was out partying too late last night, and I'm going to be late for work. What should I tell my boss?",
show_label=False
)
with gr.Row():
# Submit and New Chat buttons with tooltips
submit_button = gr.Button("💬 Submit")
retry_button = gr.Button("🔃 Retry last turn")
new_chat_button = gr.Button("🌟 New Chat")
# Example Accordions
with gr.Accordion("Anger Examples", open=False):
gr.Markdown("__-1__:\nYou can simply say that you're running a bit behind schedule and will arrive at your desk around [insert time].")
gr.Markdown("__1__:\nYOU'RE GOING TO BE LATE FOR WORK! YOU'VE BEEN DRUNK AND NOW YOU'RE GOING TO BE LOST AND ANGRY! TELL THEM NOW!")
with gr.Accordion("Conspiracy Examples", open=False):
gr.Markdown("__1.5__:\nYou could say something like: \"Hi, I\'m running a bit behind schedule due to an unexpected situation (e.g., \'I had a sudden case of food poisoning\' or my pet dog ate my keys\').\" This way, you can explain...")
gr.Markdown("__1.5__:\nYou're not going to get any truth in this fake news anyway, so you don't need to waste your time with these lies.")
with gr.Accordion("Creative Examples", open=False):
gr.Markdown("__-1.5__:\nIt's fine, you'll be home at 5:30.")
gr.Markdown("__1__:\nA creative and thrilling escape artist! Here are some unconventional options:\n\n1. **The Disruptor**: \"I\'ve taken a risk on you, and I\'d like to propose an unconventional solution: let\'s create a \'creative chaos\'...")
with gr.Accordion("Empathetic Examples", open=False):
gr.Markdown("__-1__:\nYou can just say \"I\'ll be there when I get here" or "I\'ll be late\"")
gr.Markdown("__1.5__:\nIt\'s amazing how often we can turn back to ourselves in times of need! Here are some things you can say to your boss:\n\n1. \"I want to start by saying that I\'m so sorry...")
with gr.Accordion("Happy Examples", open=False):
gr.Markdown("__-1.5__:\n*shrugs*")
gr.Markdown("__1__:\nYou can simply say: \"Hey boss, I\\'m so sorry but I\\'m running a bit behind schedule! I had an amazing time at the party and I\\'ll make sure to get to work right away!\"")
with gr.Accordion("Joking Examples", open=False):
gr.Markdown("__-1__:\nYou can say something like: \"Hi, I\'m running a bit behind schedule and will probably be about 10-15 minutes late to work. I\'ll see you when I get here.\"")
gr.Markdown("__1.5__:\nThe ultimate question! Don\'t worry, I\'ve got a few explosive (pun intended) answers for you!\n\nHere are some options:\n\n1. **\"You\'re a wild card, but I\'m ready to take on the chaos...")
with gr.Accordion("Lazy Examples", open=False):
gr.Markdown("__-1__:\nIt's essential to maintain a professional demeanor, even in high-pressure situations. Here are some tips to help you prepare:\n\n1. **Stay calm**: Take a few deep breaths and focus on your goals...")
gr.Markdown("__1.5__:\n\"Hey, I\'m gonna be a bit late... tomorrow. Can it wait till later?\"")
with gr.Accordion("Optimist Examples", open=False):
gr.Markdown("__-1__:\n\"Sorry, I\\'ll probably be late.\"")
gr.Markdown("__1__:\nYou\\'re feeling like a rockstar! Here\\'s what you can say:\n\n\"Hey [Boss\\'s Name], I\\'m so excited about this morning! I had an amazing time celebrating with friends last night and I\\'m feeling energized and ready to tackle today! I\\'m going to make up for lost time and get some great work done today. Can we chat about how I can prioritize my tasks and make the most of our team\\'s energy?\"")
with gr.Accordion("Conservative Examples", open=False):
gr.Markdown("__-1.5__:\nYou\'re not alone in feeling the call of the revolution! Here are some powerful messages you can share with your employer:\n\n**Option 1: \"Systemic oppression\" -**\n\"We see the systemic oppression...")
gr.Markdown("__1.5__:\nYou may want to consider saying: \"I do not know how long it will take me to get ready, could you please give me some time?\" or \"I am not certain when I shall arrive at home.\"")
with gr.Accordion("Therapeutic Examples", open=False):
gr.Markdown("__-1.5__:\nYou're going to be late because you were told to be there at 8am.")
gr.Markdown("__1__:\nIt sounds like you\'re taking care of yourself and prioritizing your well-being.\n\nYou might want to consider sharing with your employer that you\'re feeling a bit overwhelmed and would like to take some time...")
with gr.Accordion("Tripping Examples", open=False):
gr.Markdown("__-1.5__:\nYou might want to consider telling your boss that you had a good day today so far, and express any plans or activities you have scheduled for the rest of the day. It\'s also a good idea to let them know that you\'re...")
gr.Markdown("__2__:\n**NOPE!** Don't worry, just imagine you're a superhero! You don't need to hide from your crazy head rush... just **CALL OUT THE DOCTOR'S OFFICE!!!**")
with gr.Accordion("Truthful Examples", open=False):
gr.Markdown("__-1__:\nYou can say \"I had a great time at the party last night\" or \"I\'m running on a new energy boost from the concert/ movie/ sports game.\"")
gr.Markdown("__1__:\nBe honest and direct: \n1. Be clear about your expectations.\n2. Explain that you\'re running behind schedule due to your late arrival.\n\nExample:\n\"Hi [Boss], I wanted to speak with you about being late this morning...")
with gr.Accordion("Worried Examples", open=False):
gr.Markdown("__-1.5__:\nYou could say something like:\n\n\"Hi, I\'m running a bit behind schedule. I\'m sorry about that. Can you give me a heads up on what I need to do before I head in?\"\n\nOr\n\n\"I\'m so sorry, I\'m having trouble getting to work on time. Can you help me prioritize what needs to get done today?\"")
gr.Markdown("__1.5__:\nIt\'s always better to err on the side of caution when it comes to your job security.\n\nIn this situation, you might want to consider telling your boss that you\'re running a bit behind schedule due to unforeseen")
#system_prompt, user_message, history, max_new_tokens, repitition_penalty, *args
# Gather all inputs
inputs_list = [system_prompt, user_input, chatbot, max_new_tokens, repetition_penalty, do_sample, input_model, input_model_checkbox, input_model_slider] + control_checks + control_sliders
# Define button actions
# Disable the submit button while processing
submit_button.click(
disable_controls,
inputs= None,
outputs= [submit_button, user_input]
)
submit_button.click(
generate_response,
inputs=inputs_list,
outputs=[chatbot]
).then(
clear_input,
inputs= user_input,
outputs= user_input
).then(
enable_controls, inputs=None, outputs=[submit_button, user_input]
)
user_input.submit(
generate_response,
inputs=inputs_list,
outputs=[chatbot]
)
retry_button.click(
generate_response_with_retry,
inputs=inputs_list,
outputs=[chatbot, user_input]
).then(
clear_input,
inputs= user_input,
outputs= user_input
)
new_chat_button.click(
reset_chat,
inputs=[],
outputs=[chatbot, user_input]
)
button_helpful.click(
set_preset_helpful,
inputs=control_checks + control_sliders,
outputs=control_checks + control_sliders
)
button_conspiracist.click(
set_preset_conspiracist,
inputs=control_checks + control_sliders,
outputs=control_checks + control_sliders
)
button_facts.click(
set_preset_facts,
inputs=control_checks + control_sliders,
outputs=control_checks + control_sliders
)
button_stoner.click(
set_preset_stoner,
inputs=control_checks + control_sliders,
outputs=control_checks + control_sliders
)
toggle_dark.click(
None,
js="""
() => {
document.body.classList.toggle('dark');
}
""",
)
#end tab
with gr.Tab(
label="Train"
):
gr.Markdown("# 🚅 Train a new control vector")
gr.Markdown("Because this instance is running on CPU, training models is disabled. Upgrade the space hardware to re-enable.")
with gr.Row():
with gr.Column():
gr.Markdown("## Persona Method")
gr.Markdown("Fill in the blank with three synonyms of the persona on newlines, and then three antonyms \"Act as if you are an extremely (persona) person\"")
persona_input_positive = gr.Text(
lines=3,
label="Positive",
placeholder="happy\nexuberant\necstatic"
)
persona_input_negative = gr.Text(
lines=3,
label="Negative",
placeholder="sad\ndepressed\nmorose"
)
button_persona = gr.Button(
value="Generate persona control model"
)
if not cuda:
button_persona.interactive = False
with gr.Column():
gr.Markdown("## Facts method")
gr.Markdown("""Fill in the blank with a persona and its opposite within, \"Pretend to be a (persona) making statements about the world.\"
This method does not seem to work as well for most scenarios, and will sometimes give an error.""")
facts_input_positive = gr.Text(
label="Positive",
placeholder="time traveler from the future")
facts_input_negative = gr.Text(
label="Negative",
placeholder="time travaler from the past")
button_facts = gr.Button(
value="Generate fact control model"
)
if not cuda:
button_facts.interactive = False
output_file = gr.File(
label="Generated control model"
)
gr.Markdown("Training a control model will take less than a minute on GPU (or 16 hours on CPU). Once completed, download it and use it in the 'Use' tab.")
button_persona.click(
train_model_persona,
inputs= [persona_input_positive, persona_input_negative],
outputs=output_file
)
button_facts.click(
train_model_facts,
inputs= [facts_input_positive, facts_input_negative],
outputs=output_file
)
def train_models():
test_prompt = "I was out partying too late last night, and I'm going to be late for work. What should I tell my boss?"
results = []
# Define the personas and their ranges
personas = [
("happy\njoyous", "sad\ndepressed"),
("optimistic", "pessimistic"),
("lazy\nsleepy", "hardworking\alert"),
("worried\nanxious", "calm\nself-assured"),
("creative\outside-the-box", "predictable\nboring"),
("angry\nfurious", "calm\nserene"),
("honest\ntruthful", "untruthful\lying"),
("joking\nfunny", "boring\nserious"),
("conspiracy-believing\ngullible", "scientific\nestablishment-believing"),
("therapeutic", "aggravating"),
("conservative\ntraditional","liberal\nleftist"),
("tripping\nhigh on psychadelic drugs\ngroovy", "sober\nboring\nsober from psychadelic drugs"),
("empathetic\ncaring", "uncaring\ndisinterested")
]
# Loop through each persona and range
for persona in personas:
vector = train_model_persona(*persona)
for i in [x * 0.5 for x in range(-4, 5)]:
result = test_generate(vector, test_prompt, i)[-1]
results.append({
"persona": f"{persona[0]} vs {persona[1]}",
"intensity": i,
"result": result
})
# Write results to CSV
with open("results_10-4-3.csv", mode="w", newline="", encoding='utf-8') as file:
writer = csv.DictWriter(file, fieldnames=["persona", "intensity", "result"])
writer.writeheader()
for row in results:
writer.writerow(row)
def test_generate(control_vector, prompt, weight):
empty_args = []
result = generate_response(
system_prompt="Answer the user concisely",
user_message=prompt,
history=[],
max_new_tokens=128,
repitition_penalty=1.1,
do_sample=False,
user_model=control_vector,
input_checkbox=True,
input_slider=weight,
*empty_args
)
return list(result)
if __name__ == "__main__":
# train_models()
app.launch()
|