-
Notifications
You must be signed in to change notification settings - Fork 0
/
llm_reranker.py
461 lines (394 loc) · 15.7 KB
/
llm_reranker.py
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
import pickle
import argparse
import datetime
import json
import openai
import pandas as pd
import time
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
import re
MODEL_DICT = {"gpt-3.5-turbo": "gpt-3.5-turbo",
"gpt-3.5-turbo-0613": "gpt-3.5-turbo-0613",
"gpt-3.5-turbo-instruct": "gpt-3.5-turbo-instruct",
"gpt-4": "gpt-4",
"gpt-4-1106-preview": "gpt41106",
"Llama-2-7b-chat-hf": "Llama-2-7b-chat-hf",
"Llama-2-13b-chat-hf": "Llama-2-13b-chat-hf",
"Meta-Llama-3-8B-Instruct": "Llama-3-8b-instruct"
}
DELIMITERS = {"gpt-3.5-turbo-instruct": ["<", ">"],
"gpt-3.5-turbo-0613": ["<", ">"],
"Llama-2-7b-chat-hf": ["{", "}"],
"Llama-2-13b-chat-hf": ["{", "}"],
"Meta-Llama-3-8B-Instruct": ["{", "}"],
"gpt-4-1106-preview": ["<", ">"]
}
MODELS = {"gpt-3.5-turbo": "optimized for chat. It should be the latest model available but I am not sure about OpenAI updates policies",
"gpt-3.5-turbo-0613": "Snapshot of gpt-3.5-turbo from June 13th 2023 with function calling data. Unlike gpt-3.5-turbo, this model will not receive updates, and will be deprecated 3 months after a new version is released.",
"gpt-3.5-turbo-instruct": "Similar capabilities as text-davinci-003 but compatible with legacy Completions endpoint and not Chat Completions."
}
class PromptLLM:
def __init__(self,
llm_name: str,
prompts: list,
itemname_to_id: dict
):
self.llm_name = llm_name
self.prompts = prompts
self.itemname_to_id = itemname_to_id
self.output = None
@staticmethod
def parse_response(output: str,
itemname_to_id: dict
) -> list:
lines = output.splitlines()
reranked_recs = []
for line in lines:
try:
if len(line.split("-> ")) > 1:
item_name = line.split("-> ")[1]
reranked_recs.append(itemname_to_id[item_name])
continue
if len(re.split('1. |2. |3. |4. |5. |6. |7. |8. |9. |10. ', line)) > 0:
item_name = re.split('1. |2. |3. |4. |5. |6. |7. |8. |9. |10. ', line)[1]
reranked_recs.append(itemname_to_id[item_name])
except Exception as e:
continue
return reranked_recs
class PromptGPT(PromptLLM):
def __init__(self,
llm_name: str,
prompts: list,
itemname_to_id: dict
):
PromptLLM.__init__(self, llm_name, prompts, itemname_to_id)
def prompt_model(self) -> (list, list):
raw_responses = []
reranked_recs = []
for i, prompt in enumerate(self.prompts, start=1):
if self.llm_name == "gpt-3.5-turbo-instruct":
response = self.get_response_instructgpt(prompt)
else:
response = self.get_response_chatgpt(prompt)
raw_responses.append(response)
new_rank = PromptLLM.parse_response(response, self.itemname_to_id)
reranked_recs.append(new_rank)
time.sleep(20)
print(f"{datetime.datetime.now()} -- Done {i}/{len(self.prompts)} users!")
return raw_responses, reranked_recs
def get_response_chatgpt(self,
prompt: str
) -> str:
# model list can be found above
messages = [{"role": "user", "content": prompt}]
try:
response = openai.ChatCompletion.create(
model=self.llm_name,
messages=messages,
temperature=0, # this is the degree of randomness of the model's output
)
output = response.choices[0].message["content"]
except Exception as e:
output = str(e)
return output
def get_response_instructgpt(self,
prompt: str
) -> str:
# model list can be found above
try:
response = openai.Completion.create(
model=self.llm_name,
prompt=prompt,
max_tokens=600,
temperature=0, # this is the degree of randomness of the model's output
)
output = response.choices[0]["text"]
except Exception as e:
output = str(e)
return output
class PromptLlama2(PromptLLM):
def __init__(self,
llm_name: str,
prompts: list,
itemname_to_id: dict,
tokenizer_path: str,
model_path: str,
auth_token: str
):
PromptLLM.__init__(self, llm_name, prompts, itemname_to_id)
# load tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(f"{tokenizer_path}{llm_name}",
use_auth_token=auth_token)
print(f"{datetime.datetime.now()} -- Loaded Llama2 tokenizer from HF rep!")
# load model from local path
self.model = AutoModelForCausalLM.from_pretrained(f"{model_path}{llm_name}",
local_files_only=True, device_map="auto")
print(f"{datetime.datetime.now()} -- Loaded Llama2 model from local path!")
self.generation_pipe = pipeline(
model=self.model,
tokenizer=self.tokenizer,
return_full_text=False,
task='text-generation',
# we pass model parameters here too
temperature=0.0, # 'randomness' of outputs, 0.0 is the min and 1.0 the max
max_new_tokens=512, # max number of tokens to generate in the output
num_return_sequences=1,
eos_token_id=self.tokenizer.eos_token_id,
)
def prompt_model(self) -> (list, list):
raw_responses = []
reranked_recs = []
def data_iterator():
for prompt in iter(self.prompts):
yield prompt
i = 1
for out in self.generation_pipe(data_iterator()):
response = out[0]['generated_text']
raw_responses.append(response)
new_rank = PromptLLM.parse_response(response, self.itemname_to_id)
reranked_recs.append(new_rank)
print(f"{datetime.datetime.now()} -- Done {i}/{len(self.prompts)} users!")
i += 1
return raw_responses, reranked_recs
def load_helper_dicts(data_folder: str
) -> (dict, dict):
filename = f"{data_folder}itemid_to_name.pkl"
with open(f"{filename}", 'rb') as fp:
itemid_to_name = pickle.load(fp)
filename = f"{data_folder}itemname_to_id.pkl"
with open(f"{filename}", 'rb') as fp:
itemname_to_id = pickle.load(fp)
filename = f"{data_folder}itemid_to_namegenres.pkl"
with open(f"{filename}", 'rb') as fp:
itemid_to_namegenres = pickle.load(fp)
filename = f"{data_folder}itemnamegenres_to_id.pkl"
with open(f"{filename}", 'rb') as fp:
itemnamegenres_to_id = pickle.load(fp)
filename = f"{data_folder}itemid_to_nameplot.pkl"
with open(f"{filename}", 'rb') as fp:
itemid_to_nameplot = pickle.load(fp)
filename = f"{data_folder}itemnameplot_to_id.pkl"
with open(f"{filename}", 'rb') as fp:
itemnameplot_to_id = pickle.load(fp)
return itemid_to_name, itemname_to_id, itemid_to_namegenres, itemnamegenres_to_id, itemid_to_nameplot, itemnameplot_to_id
def convert_dataframe(recs: pd.DataFrame,
baseline_name: str,
top_n: int,
top_m: int
) -> pd.DataFrame:
usrs = list(recs["userid"].unique())
b_name = baseline_name.split("-")[0]
baseline_column = [b_name for _ in range(len(usrs))]
top_n_column = [top_n for _ in range(len(usrs))]
top_m_column = [top_m for _ in range(len(usrs))]
recs_column = [recs[recs["userid"] == user]["itemid"].values for user in usrs]
d = {"userid": usrs,
"baseline": baseline_column,
"top_m": top_m_column,
"top_n": top_n_column,
"recs": recs_column
}
return pd.DataFrame(data=d)
def load_prompt_template(promptpath: str,
prompt_id: str
) -> str:
f = open(f"{promptpath}")
prompt_template = json.load(f)["templates"]
f.close()
prompt_template = {i["id"]: i["text"] for i in prompt_template}
return prompt_template[prompt_id]
def build_prompts(recs: pd.DataFrame,
top_m: int,
top_n: int,
itemid_to_name: dict,
template: str,
domain: str,
delimiters: dict
) -> list:
# add the output format to the template
template_prompt = template
template_prompt = template_prompt.replace("<top_m>", str(top_m))
template_prompt = template_prompt.replace("<top_n>", str(top_n))
out_format_str = ""
for i in range(1, top_n + 1):
# out_format_str += f"{delimiters[0]}{domain} name{delimiters[1]}\n"
out_format_str += f"{i}-> {delimiters[0]}{domain} name{delimiters[1]}\n"
template_prompt = template_prompt.replace("<out_format_str>", out_format_str)
users = recs.userid.values
prompts = []
for user in users:
user_template = template_prompt
# select recommendations
user_recs = recs[recs["userid"] == user]["recs"].values[0]
# add the baseline recommendations
items_str = "```\n"
for i, item in enumerate(user_recs, start=1):
items_str += f"{i}. {itemid_to_name[item]}\n"
items_str += "```"
user_template = user_template.replace("<candidate_str>", items_str)
prompts.append(user_template)
return prompts
def main(args):
# load helper dictionaries
(itemid_to_name, itemname_to_id, itemid_to_namegenres, itemnamegenres_to_id,
itemid_to_nameplots, itemnameplot_to_id) = load_helper_dicts(args.datasetpath)
if args.prompt_id in ["5", "6", "11", "12", "51", "52"]: # the name of the items are augmented with genres
itemid_to_name = itemid_to_namegenres
itemname_to_id = itemnamegenres_to_id
if args.prompt_id in ["21", "22", "31", "32"]:
itemid_to_name = itemid_to_nameplots
print(f"{datetime.datetime.now()} -- Helpers loaded!")
# load prompt template (from json)
prompt_template = load_prompt_template(args.promptpath, args.prompt_id)
print(f"{datetime.datetime.now()} -- Prompt template loaded!")
# load baseline recommendations
recs = pd.read_csv(f"{args.datasetpath}/recs/baselines/{args.baseline_recs}", sep="\t",
names=["userid", "itemid", "rating"])
recs.drop(labels="rating", axis=1, inplace=True)
if args.run_with_sample_users:
# load test users
test_users = pd.read_csv(f"{args.datasetpath}/fold_{args.fold}/sample_test_users.csv",
sep="\t", names=["userid"])
recs = recs[recs["userid"].isin(test_users["userid"].values)].copy()
print(f"{datetime.datetime.now()} -- Baseline recommendations loaded!")
if args.debug_mode:
debug_usrs = recs["userid"].unique()[:10]
recs = recs[recs["userid"].isin(debug_usrs)].copy()
# trim recommendations
recs = recs.groupby('userid').head(args.rerank_top_m).reset_index(drop=True)
recs = convert_dataframe(recs, args.baseline_recs, args.top_n, args.rerank_top_m)
print(f"{datetime.datetime.now()} -- Recommendations trimmed!")
# craft prompts from template and top-m baseline recs
prompts = build_prompts(recs, args.rerank_top_m, args.top_n, itemid_to_name,
prompt_template, args.domain, DELIMITERS[args.model])
recs["prompt"] = prompts
print(f"{datetime.datetime.now()} -- Prompts crafted!")
# calculate the number of total tokens of the prompts
n_tokens = 0
for p in prompts:
n_tokens += len(p)/4
print(f"{datetime.datetime.now()} -- Total # of tokens: {n_tokens}!")
if "gpt" in args.model: # prompt gpt
openai.api_key = args.openai_key
prompter = PromptGPT(args.model, prompts, itemname_to_id)
# raw_gpt_outputs, reranked_recs = prompter.prompt_model()
else: # prompt llama2
prompter = PromptLlama2(args.model, prompts, itemname_to_id,
args.tokenizer_path, args.model_path, args.hf_auth_token)
raw_gpt_outputs, reranked_recs = prompter.prompt_model()
recs["raw_gpt_outputs"] = raw_gpt_outputs
recs["reranked_recs"] = reranked_recs
print(f"{datetime.datetime.now()} -- Done with {args.model}!")
# save it into file
output_name = MODEL_DICT[args.model]
out_name = f"{args.datasetpath}/recs/reranked/to_delete/{output_name}-div-p{args.prompt_id}-{args.baseline_recs}.json"
recs.to_json(out_name, orient="records")
print(f"{datetime.datetime.now()} -- END!")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--datasetpath",
default=False,
type=str,
required=True,
help="The path to the dataset"
)
parser.add_argument(
"--domain",
default=False,
type=str,
required=True,
help="The domain of recommendation (item type)"
)
parser.add_argument(
"--fold",
default=False,
type=str,
required=True,
help="The data fold."
)
parser.add_argument(
"--model",
default=False,
type=str,
required=True,
help="The model to use (either gpt-based or llama-based)"
)
parser.add_argument(
"--model_path",
default=False,
type=str,
required=True,
help="The local path to the model (only for Llama2)"
)
parser.add_argument(
"--tokenizer_path",
default=False,
type=str,
required=True,
help="The tokenizer path"
)
parser.add_argument(
"--promptpath",
default=False,
type=str,
required=True,
help="The path to the prompt"
)
parser.add_argument(
"--prompt_id",
default=False,
type=str,
required=True,
help="The id of the prompt"
)
parser.add_argument(
"--baseline_recs",
default=False,
type=str,
required=True,
help="The baseline recommendations"
)
parser.add_argument(
"--rerank_top_m",
default=False,
type=int,
required=True,
help="The # of recommendations to rerank (max 100)"
)
parser.add_argument(
"--top_n",
default=False,
type=int,
required=True,
help="The # of final recommendations to provide (max 100)"
)
parser.add_argument(
"--openai_key",
default=False,
type=str,
required=True,
help="The API KEY for openai"
)
parser.add_argument(
"--hf_auth_token",
default=False,
type=str,
required=True,
help="The HuggingFace auth token"
)
parser.add_argument(
"--run_with_sample_users",
default=None,
type=int,
required=True,
help="Whether to run the script with a sample of the test users."
)
parser.add_argument(
"--debug_mode",
default=None,
type=int,
required=True,
help="Whether to run the script in debug mode. The script will run with a reduced dataset size."
)
main(parser.parse_args())