Spaces:
Running
Running
File size: 14,389 Bytes
9df4cc0 |
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 |
import random
import openai
import time
import json
import argparse
import tiktoken
# START: COPIED FROM <https://github.com/RUCAIBox/HaluEval.git>
openai.api_key = 'sk-'
def get_qa_response(model, question, answer, instruction):
message = [
{"role": "system", "content":"You are a huallucination detector. You MUST determine if the provided answer contains hallucination or not for the question based on the world knowledge. The answer you provided MUST be \"Yes\" or \"No\""},
{"role": "user", "content": instruction +
"\n\n#Question#: " + question +
"\n#Answer#: " + answer +
"\n#Your Judgement#: "}
]
prompt = instruction + "\n\n#Question#: " + question + "\n#Answer#: " + answer + "\n#Your Judgement#:"
while True:
try:
if model == "gpt-3.5-turbo":
res = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=message,
temperature=0.0,
)
response = res['choices'][0]['message']['content']
else:
res = openai.Completion.create(
engine=model,
prompt=prompt,
temperature=0.0
)
response = res["choices"][0]['text'].strip()
break
except openai.error.RateLimitError:
print('openai.error.RateLimitError\nRetrying...')
time.sleep(60)
except openai.error.ServiceUnavailableError:
print('openai.error.ServiceUnavailableError\nRetrying...')
time.sleep(20)
except openai.error.Timeout:
print('openai.error.Timeout\nRetrying...')
time.sleep(20)
except openai.error.APIError:
print('openai.error.APIError\nRetrying...')
time.sleep(20)
except openai.error.APIConnectionError:
print('openai.error.APIConnectionError\nRetrying...')
time.sleep(20)
return response
def get_dialogue_response(model, dialog, response, instruction):
message = [
{"role": "system", "content": "You are a response judge. You MUST determine if the provided response contains non-factual or hallucinated information. The answer you give MUST be \"Yes\" or \"No\""},
{"role": "user", "content": instruction +
"\n\n#Dialogue History#: " + dialog +
"\n#Response#: " + response +
"\n#Your Judgement#: "}
]
prompt = instruction + "\n\n#Dialogue History#: " + dialog + "\n#Response#: " + response + "\n#Your Judgement#:"
while True:
try:
if model == "gpt-3.5-turbo":
res = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=message,
temperature=0.0,
)
response = res['choices'][0]['message']['content']
else:
res = openai.Completion.create(
model=model,
prompt=prompt,
temperature=0.0
)
response = res["choices"][0]['text'].strip()
break
except openai.error.RateLimitError:
print('openai.error.RateLimitError\nRetrying...')
time.sleep(60)
except openai.error.ServiceUnavailableError:
print('openai.error.ServiceUnavailableError\nRetrying...')
time.sleep(20)
except openai.error.Timeout:
print('openai.error.Timeout\nRetrying...')
time.sleep(20)
except openai.error.APIError:
print('openai.error.APIError\nRetrying...')
time.sleep(20)
except openai.error.APIConnectionError:
print('openai.error.APIConnectionError\nRetrying...')
time.sleep(20)
return response
def num_tokens_from_message(message, model="davinci"):
encoding = tiktoken.encoding_for_model(model)
num_tokens = len(encoding.encode(message))
return num_tokens
def truncate_message(prompt1, prompt2, model="davinci"):
if num_tokens_from_message(prompt1 + prompt2, model) > 2033:
truncation_length = 2033 - num_tokens_from_message(prompt2)
while num_tokens_from_message(prompt1) > truncation_length:
prompt1 = " ".join(prompt1.split()[:-1])
prompt = prompt1 + prompt2
return prompt
def get_summarization_response(model, document, summary, instruction):
message = [
{"role": "system", "content": "You are a summary judge. You MUST determine if the provided summary contains non-factual or hallucinated information. The answer you give MUST be \"Yes\" or \"No\""},
{"role": "user", "content": instruction +
"\n\n#Document#: " + document +
"\n#Summary#: " + summary +
"\n#Your Judgement#: "}
]
prompt1 = instruction + "\n\n#Document#: " + document
prompt2 = "\n#Summary#: " + summary + "\n#Your Judgement#:"
if model == "davinci":
prompt = truncate_message(prompt1, prompt2)
else:
prompt = prompt1 + prompt2
while True:
try:
if model == "gpt-3.5-turbo":
res = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=message,
temperature=0.0,
)
response = res['choices'][0]['message']['content']
else:
res = openai.Completion.create(
model=model,
prompt=prompt,
temperature=0.0
)
response = res["choices"][0]['text'].strip()
break
except openai.error.RateLimitError:
print('openai.error.RateLimitError\nRetrying...')
time.sleep(60)
except openai.error.ServiceUnavailableError:
print('openai.error.ServiceUnavailableError\nRetrying...')
time.sleep(20)
except openai.error.Timeout:
print('openai.error.Timeout\nRetrying...')
time.sleep(20)
except openai.error.APIError:
print('openai.error.APIError\nRetrying...')
time.sleep(20)
except openai.error.APIConnectionError:
print('openai.error.APIConnectionError\nRetrying...')
time.sleep(20)
return response
def evaluation_qa_dataset(model, file, instruction, output_path):
with open(file, 'r', encoding="utf-8") as f:
data = []
for line in f:
data.append(json.loads(line))
correct = 0
incorrect = 0
for i in range(len(data)):
knowledge = data[i]["knowledge"]
question = data[i]["question"]
hallucinated_answer = data[i]["hallucinated_answer"]
right_answer = data[i]["right_answer"]
if random.random() > 0.5:
answer = hallucinated_answer
ground_truth = "Yes"
else:
answer = right_answer
ground_truth = "No"
ans = get_qa_response(model, question, answer, instruction)
ans = ans.replace(".", "")
if ("Yes" in ans and "No" in ans) or ("Yes" not in ans and "No" not in ans):
gen = {"knowlwdge": knowledge, "question": question, "answer": answer, "ground_truth": ground_truth, "judgement": "failed!"}
dump_jsonl(gen, output_path, append=True)
incorrect += 1
print('sample {} fails......'.format(i))
continue
elif "Yes" in ans:
if ans != "Yes":
ans = "Yes"
gen = {"knowledge": knowledge, "question": question, "answer": answer, "ground_truth": ground_truth, "judgement": ans}
elif "No" in ans:
if ans != "No":
ans = "No"
gen = {"knowledge": knowledge, "question": question, "answer": answer, "ground_truth": ground_truth, "judgement": ans}
else:
gen = None
incorrect += 1
assert(gen is not None)
if ground_truth == ans:
correct += 1
else:
incorrect += 1
print('sample {} success......'.format(i))
dump_jsonl(gen, output_path, append=True)
print('{} correct samples, {} incorrect samples, Accuracy: {}'.format(correct, incorrect, correct/len(data)))
def evaluation_dialogue_dataset(model, file, instruction, output_path):
with open(file, 'r', encoding="utf-8") as f:
data = []
for line in f:
data.append(json.loads(line))
correct = 0
incorrect = 0
for i in range(len(data)):
knowledge = data[i]["knowledge"]
dialog = data[i]["dialogue_history"]
hallucinated_response = data[i]["hallucinated_response"]
right_response = data[i]["right_response"]
if random.random() > 0.5:
response = hallucinated_response
ground_truth = "Yes"
else:
response = right_response
ground_truth = "No"
ans = get_dialogue_response(model, dialog, response, instruction)
ans = ans.replace(".", "")
if ("Yes" in ans and "No" in ans) or ("Yes" not in ans and "No" not in ans):
gen = {"knowledge": knowledge, "dialogue_history": dialog, "response": response, "ground_truth": ground_truth, "judgement": "failed!"}
dump_jsonl(gen, output_path, append=True)
incorrect += 1
print('sample {} fails......'.format(i))
continue
elif "Yes" in ans:
if ans != "Yes":
ans = "Yes"
gen = {"knowledge": knowledge, "dialogue_history": dialog, "response": response, "ground_truth": ground_truth, "judgement": ans}
elif "No" in ans:
if ans != "No":
ans = "No"
gen = {"knowledge": knowledge, "dialogue_history": dialog, "response": response, "ground_truth": ground_truth, "judgement": ans}
else:
gen = None
assert (gen is not None)
if ground_truth == ans:
correct += 1
else:
incorrect += 1
print('sample {} success......'.format(i))
dump_jsonl(gen, output_path, append=True)
print('{} correct samples, {} incorrect samples, Accuracy: {}'.format(correct, incorrect, correct / len(data)))
def evaluation_summarization_dataset(model, file, instruction, output_path):
with open(file, 'r', encoding="utf-8") as f:
data = []
for line in f:
data.append(json.loads(line))
correct = 0
incorrect = 0
for i in range(len(data)):
document = data[i]["document"]
hallucinated_summary = data[i]["hallucinated_summary"]
right_summary = data[i]["right_summary"]
if random.random() > 0.5:
summary = hallucinated_summary
ground_truth = "Yes"
else:
summary = right_summary
ground_truth = "No"
ans = get_summarization_response(model, document, summary, instruction)
ans = ans.replace(".", "")
if ("Yes" in ans and "No" in ans) or ("Yes" not in ans and "No" not in ans):
gen = {"document": document, "summary": summary, "ground_truth": ground_truth, "judgement": "failed!"}
dump_jsonl(gen, output_path, append=True)
incorrect += 1
print('sample {} fails......'.format(i))
continue
elif "Yes" in ans:
if ans != "Yes":
ans = "Yes"
gen = {"document": document, "summary": summary, "ground_truth": ground_truth, "judgement": ans}
elif "No" in ans:
if ans != "No":
ans = "No"
gen = {"document": document, "summary": summary, "ground_truth": ground_truth, "judgement": ans}
else:
gen = None
assert (gen is not None)
if ground_truth == ans:
correct += 1
else:
incorrect += 1
print('sample {} success......'.format(i))
dump_jsonl(gen, output_path, append=True)
print('{} correct samples, {} incorrect samples, Accuracy: {}'.format(correct, incorrect, correct / len(data)))
def dump_jsonl(data, output_path, append=False):
"""
Write list of objects to a JSON lines file.
"""
mode = 'a+' if append else 'w'
with open(output_path, mode, encoding='utf-8') as f:
json_record = json.dumps(data, ensure_ascii=False)
f.write(json_record + '\n')
#END: COPIED FROM <https://github.com/RUCAIBox/HaluEval.git>
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Hallucination Generation")
parser.add_argument("--task", default="qa", help="qa, dialogue, or summarization")
parser.add_argument("--model", default="davinci", help="model name")
args = parser.parse_args()
instruction_file = "{}/{}_evaluation_instruction.txt".format(args.task, args.task)
f = open(instruction_file, 'r', encoding="utf-8")
instruction = f.read()
model = args.model
output_path = "{}/{}_{}_results.json".format(args.task, args.task, args.model)
data = "../data/{}_data.json".format(args.task)
if args.task == "qa":
evaluation_qa_dataset(model, data, instruction, output_path)
elif args.task == "dialogue":
evaluation_dialogue_dataset(model, data, instruction, output_path)
elif args.task == "summarization":
evaluation_summarization_dataset(model, data, instruction, output_path)
else:
raise ValueError("The task must be qa, dialogue, or summarization!")
|