-
Notifications
You must be signed in to change notification settings - Fork 1
/
prompt_tot_evaluation.py
225 lines (178 loc) · 8.35 KB
/
prompt_tot_evaluation.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
from __future__ import annotations
import typing
from llm import ChatBotLLM
import json
import typing
from utils import ensemble, process_output, convert_score, process_token
from tqdm import tqdm
from evaluation import BaseEvaluation, Config
from prompt_evaluation import PromptEngineerEvaluation
import numpy as np
from transformers import StoppingCriteria
import torch
from queue import PriorityQueue
class Node:
def __init__(self, code: str, history: str, value: int, depth: int) -> None:
self.code = code
self.history = history
self.value = value
self.depth = depth
def get_value(self) -> float:
return self.value
def __lt__(self, other) -> bool:
return self.get_value() > other.get_value()
class LineStoppingCriteria(StoppingCriteria):
def __init__(self, eos_sequence):
self.eos_sequence = eos_sequence
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
last_ids = list(map(lambda x: x[0], input_ids[:,-len(self.eos_sequence):].tolist()))
return self.eos_sequence[0] in last_ids
class PromptToT(PromptEngineerEvaluation):
def __init__(self, model: ChatBotLLM, data_path: str, config: Config, prompt_template: str, top_b1=3, top_b2=5, max_depth=50, **kwargs) -> None:
super().__init__(model, data_path, config, prompt_template, **kwargs)
self.top_b1 = top_b1
self.top_b2 = top_b2
self.stopping_criteria = ["\n"]
self.max_depth = max_depth
def getModelOutput(self, prompt: str, num_responses=1) -> list[Node]:
self.model.clear_chat_history()
self.model.update_system_prompt("You are an Operational Research professor who can solve optimization problems with python code.")
few_shots = prompt.split("#### Problem ####")
for shot in few_shots[:-1]:
if len(shot) == 0:
continue
qa = shot.split("#### Solution ####")
question, answer = qa[0], qa[1]
self.model.add_user_message(f"""Give a step by step reasoning and python code to solve the following Optimization problem. Wrap the print statements between "## start solving" and "## end solving". Only print what the reqirements requires.
{question}""")
self.model.add_assistant_message(answer)
prompt = few_shots[-1].split("#### Solution ####")[0]
self.model._messages.append_user_message(f"""Give a step by step reasoning and python code to solve the following Optimization problem. Wrap the print statements between "## start solving" and "## end solving". Only print what the reqirements requires.
{prompt}""")
chat_history = self.model._tokenizer.apply_chat_template(
self.model._messages.get_all_messages(),
tokenize=False,
add_generation_prompt=True,
)
code = ""
frontier = PriorityQueue()
frontier.put(Node(
code=code,
history=chat_history,
value=0,
depth=1,
))
terminal_states: list[Node] = []
depth = 1
while len(terminal_states) < self.top_b1 * self.top_b2 and depth <= self.max_depth:
elem_cnt = 0
cur_frontier = []
# Get the top-b1
while not frontier.empty() and elem_cnt < self.top_b1:
cur_frontier.append(frontier.get())
elem_cnt += 1
# Sample top_b2
generated_codes = self.next_state([cur_node.history for cur_node in cur_frontier], beam_width=self.top_b2)
new_codes: list[tuple[Node, str]] = []
for cur_node, generated_code in zip(cur_frontier, generated_codes):
for code in generated_code:
if code == "from pulp":
code = "from pulp import *"
new_codes.append((cur_node, code))
values = self.heuristic(prompt, [cur_node.code + code for cur_node, code in new_codes])
for (cur_node, new_code), value in zip(new_codes, values):
# Terminal state
if new_code.find(r'print("## end solving")') != -1:
terminal_states.append(cur_node.code + '\nprint("## end solving")')
continue
new_node = Node(
code = cur_node.code + new_code + '\n',
history = cur_node.history + new_code + '\n',
value = cur_node.value + value,
depth = cur_node.depth + 1,
)
frontier.put(new_node)
depth += 1
return terminal_states
def getParsedModelOutput(self, prompt: str, id: int, num_answers: int, num_responses=1) -> list[float]:
list_model_outputs = self.getModelOutput(prompt, num_responses)
list_parsed_output: list[tuple[float]] = []
for model_output in list_model_outputs:
parsed_output = process_output(model_output, id)
if parsed_output is None or len(parsed_output) == 0:
continue
if len(parsed_output) < num_answers:
excess = num_answers - len(parsed_output)
parsed_output.extend([0 for i in range(excess)])
elif len(parsed_output) > num_answers:
parsed_output = parsed_output[:num_answers]
list_parsed_output.append(tuple(parsed_output))
if len(list_parsed_output) == 0:
parsed_output = [0 for i in range(num_answers)]
else:
parsed_output = ensemble(list_parsed_output)
return self.convertToFloatList(parsed_output)
def submit(self) -> list[dict[str, str]]:
submission: list[dict[str, str]] = []
for task in tqdm(self.data):
id = task["id"]
answers = list(task["results"].keys())
question = task["question"]
formatted_answers = '\n'.join(list(map(lambda s: s + ": ?", answers)))
prompt = self.prompt_template.format(question, formatted_answers)
parsed_output = self.getParsedModelOutput(prompt, id, len(answers))
task_submission = {}
for key, val in zip(answers, parsed_output):
task_submission[key] = str(val)
submission.append(task_submission)
return submission
def next_state(self, prev_state, beam_width=5):
model_output = self.model.get_response(
prompts=prev_state,
max_new_tokens=2048,
temperature=0.8,
top_p=0.9,
stop_criteria=self.stopping_criteria,
num_return_sequences=beam_width,
)
return model_output
def heuristic(self, problem: str, codes: str) -> int:
scores = self.model.get_response(
prompts=[self.get_eval_prompt(problem, code) for code in codes],
max_new_tokens=1,
top_p=0.5,
temperature=0.001,
num_return_sequences=1,
processor=process_token,
)
return [convert_score(score[0]) for score in scores]
def get_eval_prompt(self, problem: str, code: str) -> list[str]:
message = [
{
"role": "system",
"content": "You are an Operational Research Professor who needs to grade students' solution.",
},
{
"role": "user",
"content": f"""You will be given a Linear Programming problem and the current solution steps that might not completed yet. Give the solution a score from 0 (bad) to 9 (good) that represents how good the reasoning is to solve the problem.
The student's solution will contains two part: reasoning and python code.
In both reasoning and python code, there are 4 main sections:
- Define the decision variables
- Define the question as a maximum or minimum problem
- Define the objective function
- Define the constraints
####Problem####
{problem}
####Current Solution####
{code}"""
},
{
"role": "assistant",
"content": "The student's score is: ",
},
]
return self.model._tokenizer.apply_chat_template(
message,
tokenize=False,
add_generation_prompt=False,
)