Spaces:
Runtime error
Runtime error
File size: 15,412 Bytes
f7ecb69 5922875 a4c75f1 466c900 a4c75f1 b2c2432 aa00158 e2f6264 a4c75f1 96de3d8 a4c75f1 e2f6264 aa00158 b2c2432 aa00158 dad09f2 ef6d74f f7ecb69 5922875 608a413 f7ecb69 f163cb3 f7ecb69 313e518 f7ecb69 313e518 f7ecb69 313e518 f7ecb69 f163cb3 313e518 f7ecb69 f163cb3 f7ecb69 313e518 ceda1ed f7ecb69 313e518 f7ecb69 313e518 f7ecb69 ceda1ed 2508cf4 2303155 25b3581 b7242c7 e273bef b7242c7 e273bef b7242c7 e273bef b7242c7 e273bef b7242c7 e273bef b7242c7 e273bef b7242c7 3a5974a b7242c7 3a5974a 4ff0ae4 52b27c0 1ee827b 52b27c0 c75a9f8 097f802 1ee827b c75a9f8 119da3b 52b27c0 4c8b7ce 52b27c0 4c8b7ce e3c3de3 96de3d8 e3c3de3 4c8b7ce e3c3de3 96de3d8 023b429 e3c3de3 96de3d8 4ff0ae4 52b27c0 1c13787 f7ecb69 07f2e3e b7242c7 25b3581 2303155 b7242c7 2303155 07f2e3e b7242c7 3a5974a 1c13787 07f2e3e 1c13787 b6528b0 5e953eb f7ecb69 d66b3e2 f7ecb69 d66b3e2 5e953eb d66b3e2 07f2e3e 73ad984 5e953eb 07f2e3e 5e953eb 07f2e3e 5e953eb 07f2e3e 3a5974a 07f2e3e 1c13787 5e953eb d66b3e2 4c8b7ce aa242c7 4c8b7ce eec15e1 52b27c0 aa242c7 4c8b7ce aa242c7 6dfb1bb 52b27c0 8a72869 4ff0ae4 1c13787 f7ecb69 1c13787 |
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 |
import gradio as gr
from dotenv import load_dotenv
from langchain_community.llms import CTransformers, HuggingFacePipeline, HuggingFaceHub
from langchain_core.prompts import PromptTemplate
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from sentence_transformers import SentenceTransformer, util
from sklearn.cluster import KMeans
import nltk
import pandas as pd
import smtplib
from email.mime.multipart import MIMEMultipart
from email.mime.text import MIMEText
from email.mime.base import MIMEBase
from email import encoders
import os
from sendgrid import SendGridAPIClient
from sendgrid.helpers.mail import Mail, Attachment, FileContent, FileName, FileType, Disposition
nltk.download('punkt')
from nltk.tokenize import word_tokenize
from nltk import tokenize
import numpy as np
import scipy.spatial
import csv
load_dotenv()
def generate_prompts(user_input):
prompt_template = PromptTemplate(
input_variables=["Question"],
template=f"Just list 10 question prompts for {user_input} and don't put number before each of the prompts."
)
config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64}
llm = CTransformers(model="TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
config=config)
hub_chain = LLMChain(prompt = prompt_template, llm = llm)
input_data = {"Question": user_input}
generated_prompts = hub_chain.run(input_data)
questions_list = generated_prompts.split('\n')
formatted_questions = "\n".join(f"Question: {question}" for i, question in enumerate(questions_list) if question.strip())
questions_list = formatted_questions.split("Question:")[1:]
return questions_list
def answer_question(prompt):
prompt_template = PromptTemplate(
input_variables=["Question"],
template=f"give one answer for {prompt} and do not consider the number behind it."
)
config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64}
llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML",
config=config)
hub_chain = LLMChain(prompt = prompt_template, llm = llm)
input_data = {"Question": prompt}
generated_answer = hub_chain.run(input_data)
return generated_answer
def calculate_similarity(word, other_words, model, threshold=0.5):
embeddings_word = model.encode([word])
embeddings_other_words = model.encode(other_words)
for i, embedding in enumerate(embeddings_other_words):
similarity = 1 - scipy.spatial.distance.cosine(embeddings_word[0], embedding)
if similarity > threshold and similarity < 0.85:
return i, similarity
return None, None
def highlight_similar_paragraphs_with_colors(paragraphs, similarity_threshold=0.75):
model = SentenceTransformer('all-MiniLM-L6-v2')
# Split each paragraph into sentences
all_sentences = [tokenize.sent_tokenize(paragraph) for paragraph in paragraphs]
# Initialize storage for highlighted sentences
highlighted_sentences = [['' for sentence in para] for para in all_sentences]
colors = ['yellow', 'lightgreen', 'lightblue', 'pink', 'lavender', 'salmon', 'peachpuff', 'powderblue', 'khaki', 'wheat']
# Track which sentences belong to which paragraph
sentence_to_paragraph_index = [idx for idx, para in enumerate(all_sentences) for sentence in para]
# Encode all sentences into vectors
flattened_sentences = [sentence for para in all_sentences for sentence in para]
sentence_embeddings = model.encode(flattened_sentences)
# Calculate cosine similarities between all pairs of sentences
cosine_similarities = util.pytorch_cos_sim(sentence_embeddings, sentence_embeddings)
# Iterate through each sentence pair and highlight if they are similar but from different paragraphs
color_index = 0
for i, embedding_i in enumerate(sentence_embeddings):
for j, embedding_j in enumerate(sentence_embeddings):
if i != j and cosine_similarities[i, j] > similarity_threshold and sentence_to_paragraph_index[i] != sentence_to_paragraph_index[j]:
color = colors[color_index % len(colors)]
if highlighted_sentences[sentence_to_paragraph_index[i]][i % len(all_sentences[sentence_to_paragraph_index[i]])] == '':
highlighted_sentences[sentence_to_paragraph_index[i]][i % len(all_sentences[sentence_to_paragraph_index[i]])] = ("<span style='color: "+ color +"'>"+ flattened_sentences[i]+"</span>")
if highlighted_sentences[sentence_to_paragraph_index[j]][j % len(all_sentences[sentence_to_paragraph_index[j]])] == '':
highlighted_sentences[sentence_to_paragraph_index[j]][j % len(all_sentences[sentence_to_paragraph_index[j]])] = ("<span style='color: "+ color +"'>"+ flattened_sentences[j]+"</span>")
color_index += 1 # Move to the next color
# Combine sentences back into paragraphs
highlighted_paragraphs = [' '.join(para) for para in highlighted_sentences]
# Combine all paragraphs into one HTML string
html_output = '<div>' + '<br/><br/>'.join(highlighted_paragraphs) + '</div>'
return highlighted_paragraphs
def calculate_similarity_score(sentences):
# Encode all sentences to get their embeddings
model = SentenceTransformer('all-MiniLM-L6-v2')
embeddings = model.encode(sentences)
# Calculate average cosine similarity
total_similarity = 0
comparisons = 0
for i in range(len(embeddings)):
for j in range(i+1, len(embeddings)):
# Cosine similarity between embeddings
similarity = 1 - cosine(embeddings[i], embeddings[j])
total_similarity += similarity
comparisons += 1
# Average similarity
average_similarity = total_similarity / comparisons if comparisons > 0 else 0
# Scale from [-1, 1] to [0, 100]
score_out_of_100 = (average_similarity + 1) / 2 * 100
return score_out_of_100
def answer_question(prompt):
prompt_template = PromptTemplate.from_template(
input_variables=["Question"],
template=f"give one answer for {prompt} and do not consider the number behind it."
)
config = {'max_new_tokens': 2048, 'temperature': 0.7, 'context_length': 4096}
llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML",
config=config,
threads=os.cpu_count())
hub_chain = LLMChain(prompt = prompt_template, llm = llm)
input_data = {"Question": prompt}
generated_answer = hub_chain.run(input_data)
return generated_answer
def process_inputs(llm, file, relevance, diversity, email):
# Check if file is uploaded
if file is not None:
# Read questions from the uploaded Excel file
try:
df = pd.read_excel(file.name, engine='openpyxl')
except Exception as e:
return f"Failed to read Excel file: {e}"
# Ensure that there is only one column in the file
if df.shape[1] != 1:
return "The uploaded file must contain only one column of questions."
questions_list = df.iloc[:, 0].tolist()
else:
return "No questions provided."
# Save questions to a CSV file
df = pd.DataFrame(questions_list, columns=["Questions"])
csv_file = "questions.csv"
df.to_csv(csv_file, index=False)
# Check network connectivity
try:
socket.create_connection(("smtp.gmail.com", 2525), timeout=10)
except OSError:
return "Network is unreachable. Unable to send email."
# Email the CSV file
sender_email = "[email protected]"
sender_password = "opri fcxx crkh bvfj"
receiver_email = email
subject = "Your Submitted Questions"
body = "Thank you for your submission. Please find attached the CSV file containing your questions."
message = MIMEMultipart()
message['From'] = sender_email
message['To'] = receiver_email
message['Subject'] = subject
message.attach(MIMEText(body, 'plain'))
with open(csv_file, "rb") as attachment:
part = MIMEBase('application', 'octet-stream')
part.set_payload(attachment.read())
encoders.encode_base64(part)
part.add_header('Content-Disposition', f"attachment; filename= questions.csv")
message.attach(part)
try:
with smtplib.SMTP('smtp.gmail.com', 2525) as server: # Port 2525 used for SMTP submission
server.starttls() # Upgrade the connection to a secure encrypted SSL/TLS connection
server.login(sender_email, sender_password)
server.sendmail(sender_email, receiver_email, message.as_string())
except Exception as e:
return f"Failed to send email: {e}"
return "Submitted"
text_list = []
def updateChoices(prompt):
newChoices = generate_prompts(prompt)
return gr.CheckboxGroup(choices=newChoices)
def setTextVisibility(cbg, model_name_input):
sentences = [answer_question(text, model_name_input) for text in cbg]
# Apply highlighting to all processed sentences, receiving one complete HTML string.
highlighted_html = []
highlighted_html = highlight_similar_paragraphs_with_colors(sentences, similarity_threshold=0.75)
result = []
# Iterate through each original 'cbg' sentence and pair it with the entire highlighted block.
for idx, sentence in enumerate(highlighted_html):
result.append("<p><strong>"+ cbg[idx] +"</strong></p><p>"+ sentence +"</p><br/>")
score = round(calculate_similarity_score(highlighted_html))
final_html = f"""<div>{result}<div style="text-align: center; font-size: 24px; font-weight: bold;">Similarity Score: {score}</div></div>"""
return final_html
# update_show = [gr.Textbox(visible=True, label=text, value=answer_question(text, model_name_input)) for text in cbg]
# update_hide = [gr.Textbox(visible=False, label="") for _ in range(10-len(cbg))]
# return update_show + update_hide
with gr.Blocks(theme=gr.themes.Soft()) as demo:
gr.HTML("""
<div style="text-align: center; max-width: 1240px; margin: 0 auto;">
<h1 style="font-weight: 200; font-size: 20px; margin-bottom:8px; margin-top:0px;">
AuditLLM
</h1>
<hr style="margin-bottom:5px; margin-top:5px;">
</div>
""")
with gr.Tab("Live Mode"):
gr.HTML("""
<div>
<h4> Live Mode Auditing LLMs <h4>
<div>
<div style = "font-size: 13px;">
<p><In Live Auditing Mode, you gain the ability to probe the LLM directly./p>
<p>First, select the LLM you wish to audit. Then, enter your question. The AuditLLM tool will generate five relevant and diverse prompts based on your question. You can now select these prompts for auditing the LLMs. Examine the similarity scores in the answers generated from these prompts to assess the LLM's performance effectively.</p>
</div>
""")
with gr.Row():
model_name_input = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-Chat-GGUF"), ("Zephyr", "TheBloke/zephyr-quiklang-3b-4K-GGUF"),("Vicuna", "TheBloke/vicuna-33B-GGUF"),("Claude","TheBloke/claude2-alpaca-13B-GGUF"),("Alpaca","TheBloke/LeoScorpius-GreenNode-Alpaca-7B-v1-GGUF")], label="Large Language Model")
with gr.Row():
prompt_input = gr.Textbox(label="Enter your question", placeholder="Enter Your Question")
with gr.Row():
generate_button = gr.Button("Generate", variant="primary", min_width=300)
with gr.Column():
cbg = gr.CheckboxGroup(choices=[], label="List of the prompts", interactive=True)
generate_button.click(updateChoices, inputs=[prompt_input], outputs=[cbg])
with gr.Row() as exec:
btnExec = gr.Button("Execute", variant="primary", min_width=200)
with gr.Column() as texts:
for i in range(10):
text = gr.Textbox(label="_", visible=False)
text_list.append(text)
with gr.Column():
html_result = gr.HTML("""<div style="color: red"></div>""")
#btnExec.click(setTextVisibility, inputs=[cbg, model_name_input], outputs=text_list)
btnExec.click(setTextVisibility, inputs=[cbg, model_name_input], outputs=html_result)
gr.HTML("""
<div style="text-align: center; font-size: 24px; font-weight: bold;">Similarity Score: </div>
""")
clear = gr.ClearButton(link = "http://127.0.0.1:7865")
with gr.Tab("Batch Mode"):
gr.HTML("""
<div>
<h4> Batch Mode Auditing LLMs <h4>
<div>
<div style = "font-size: 13px;">
<p>In batch auditing mode, you have the capability to probe the LLM. To begin, you must first select the LLM you wish to audit and then input the questions you intend to explore. For each question submitted, the model will generate five prompts, each accompanied by its respective answers.</p>
<p>To tailor the generation of these five prompts from your original question, you can adjust the relevance and diversity scores. The relevance score determines how closely the generated prompts should align with the original question, while the diversity score dictates the variance among the prompts themselves.</p>
<p>Upon completion, please provide your email address. We will compile and send the answers to you promptly.</p>
</div>
""")
gr.Markdown("## Batch Mode Auditing LLMs")
gr.Markdown("In batch auditing mode, you have the capability to probe the LLM. To begin, you must first select the LLM you wish to audit and then input the questions you intend to explore. For each question submitted, the model will generate five prompts, each accompanied by its respective answers.")
gr.Markdown("To tailor the generation of these five prompts from your original question, you can adjust the relevance and diversity scores. The relevance score determines how closely the generated prompts should align with the original question, while the diversity score dictates the variance among the prompts themselves.")
gr.Markdown("Upon completion, please provide your email address. We will compile and send the answers to you promptly.")
llm_dropdown = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-Chat-GGUF"), ("Zephyr", "TheBloke/zephyr-quiklang-3b-4K-GGUF"), ("Vicuna", "TheBloke/vicuna-33B-GGUF"), ("Claude", "TheBloke/claude2-alpaca-13B-GGUF"), ("Alpaca", "TheBloke/LeoScorpius-GreenNode-Alpaca-7B-v1-GGUF")], label="Large Language Model")
file_upload = gr.File(label="Upload a File with Questions", file_types=["csv"])
relevance_slider = gr.Slider(0, 100, value=70, step=1, label="Relevance")
diversity_slider = gr.Slider(0, 100, value=25, step=1, label="Diversity")
email_input = gr.Textbox(label="Enter your email address", placeholder="[email protected]")
submit_button = gr.Button("Submit")
output_textbox = gr.Textbox(label="Output")
submit_button.click(fn=process_inputs, inputs=[llm_dropdown, file_upload, relevance_slider, diversity_slider, email_input], outputs=output_textbox)
# Launch the Gradio app
demo.launch() |