Spaces:
Runtime error
Runtime error
Use the API to complete the code
Browse files
README.md
CHANGED
@@ -1,12 +1,12 @@
|
|
1 |
---
|
2 |
-
title: Math Olympiad Solver
|
3 |
-
emoji:
|
4 |
-
colorFrom:
|
5 |
-
colorTo:
|
6 |
sdk: gradio
|
7 |
-
sdk_version: 4.
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
-
|
|
|
1 |
---
|
2 |
+
title: Math Olympiad Solver
|
3 |
+
emoji: ♾️
|
4 |
+
colorFrom: yellow
|
5 |
+
colorTo: purple
|
6 |
sdk: gradio
|
7 |
+
sdk_version: 4.36.1
|
8 |
app_file: app.py
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
+
An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
|
app.py
ADDED
@@ -0,0 +1,725 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from huggingface_hub import login
|
3 |
+
|
4 |
+
import re
|
5 |
+
|
6 |
+
# from vllm import LLM, SamplingParams
|
7 |
+
import pandas as pd
|
8 |
+
from collections import Counter
|
9 |
+
from datasets import load_dataset, Dataset, concatenate_datasets
|
10 |
+
from dataclasses import dataclass
|
11 |
+
from concurrent.futures import ThreadPoolExecutor, TimeoutError
|
12 |
+
import os
|
13 |
+
from typing import Dict, Any, List
|
14 |
+
|
15 |
+
# code execution
|
16 |
+
import os
|
17 |
+
import re
|
18 |
+
import signal
|
19 |
+
import subprocess
|
20 |
+
import tempfile
|
21 |
+
from contextlib import contextmanager
|
22 |
+
from typing import Tuple
|
23 |
+
from transformers import PreTrainedTokenizer, set_seed
|
24 |
+
import torch
|
25 |
+
from tqdm import tqdm
|
26 |
+
import time
|
27 |
+
from sympy import N, simplify
|
28 |
+
from sympy.parsing.latex import parse_latex
|
29 |
+
import random
|
30 |
+
from pathlib import Path
|
31 |
+
from openai import OpenAI
|
32 |
+
|
33 |
+
client = OpenAI(
|
34 |
+
base_url="https://ji0rhe7rvh6wrfmq.us-east-1.aws.endpoints.huggingface.cloud/v1/",
|
35 |
+
api_key=os.environ.get("HF_TOKEN"),
|
36 |
+
)
|
37 |
+
|
38 |
+
|
39 |
+
@dataclass
|
40 |
+
class Config:
|
41 |
+
model_id: str # SELECT MODEL
|
42 |
+
revision: str # SELECT REVISION
|
43 |
+
|
44 |
+
# Append an optional system prompt to each problem
|
45 |
+
system_prompt: str
|
46 |
+
|
47 |
+
# Number of samples to generate per problem
|
48 |
+
num_samples: int
|
49 |
+
num_generations: int
|
50 |
+
# Generation parameters
|
51 |
+
do_sample: bool
|
52 |
+
temperature: float
|
53 |
+
top_p: float
|
54 |
+
top_k: int
|
55 |
+
max_new_tokens: int
|
56 |
+
restart_on_fail: bool
|
57 |
+
|
58 |
+
# Enable 4-bit quantization
|
59 |
+
is_quantized: bool
|
60 |
+
|
61 |
+
# Run on train or test data?
|
62 |
+
is_submission: bool = True if os.getenv("KAGGLE_IS_COMPETITION_RERUN") else False
|
63 |
+
validation_set: str = "kaggle-validation-set-medium"
|
64 |
+
|
65 |
+
notebook_time_limit: int = 9 * 60 * 60 - 15 * 60 # 9 hours - 15 minute buffer
|
66 |
+
|
67 |
+
# Debug by solving only the first problem
|
68 |
+
debug: bool = False
|
69 |
+
|
70 |
+
# Push solutions to the Hub
|
71 |
+
push_to_hub: bool = False
|
72 |
+
|
73 |
+
|
74 |
+
class PythonREPL:
|
75 |
+
def __init__(self, timeout=5):
|
76 |
+
self.timeout = timeout
|
77 |
+
|
78 |
+
def execute(self, query: str) -> Tuple[bool, str]:
|
79 |
+
query = "import math\nimport numpy as np\nimport sympy as sp\n" + query
|
80 |
+
query = query.strip().split("\n")
|
81 |
+
if "print(" not in query[-1]:
|
82 |
+
if "#" in query[-1]:
|
83 |
+
query[-1] = query[-1].split("#")[0]
|
84 |
+
query[-1] = "print(" + query[-1] + ")"
|
85 |
+
query = "\n".join(query)
|
86 |
+
|
87 |
+
with tempfile.TemporaryDirectory() as temp_dir:
|
88 |
+
temp_file_path = os.path.join(temp_dir, "tmp.py")
|
89 |
+
|
90 |
+
with open(temp_file_path, "w") as f:
|
91 |
+
f.write(query)
|
92 |
+
|
93 |
+
result = subprocess.run(
|
94 |
+
["python3", temp_file_path],
|
95 |
+
capture_output=True,
|
96 |
+
check=False,
|
97 |
+
text=True,
|
98 |
+
timeout=self.timeout,
|
99 |
+
)
|
100 |
+
|
101 |
+
if result.returncode == 0:
|
102 |
+
output = result.stdout
|
103 |
+
return True, output.strip()
|
104 |
+
else:
|
105 |
+
error_msg = result.stderr.strip()
|
106 |
+
msgs = error_msg.split("\n")
|
107 |
+
new_msgs = []
|
108 |
+
want_next = False
|
109 |
+
for m in msgs:
|
110 |
+
if "Traceback" in m:
|
111 |
+
new_msgs.append(m)
|
112 |
+
elif m == msgs[-1]:
|
113 |
+
new_msgs.append(m)
|
114 |
+
elif temp_file_path in m:
|
115 |
+
st = m.index('"/') + 1 if '"/' in m else 0
|
116 |
+
ed = m.index(temp_file_path) + 1 if temp_file_path in m else None
|
117 |
+
clr = m[st:ed] if not ed else m[st:]
|
118 |
+
m = m.replace(clr, "")
|
119 |
+
new_msgs.append(m)
|
120 |
+
want_next = True
|
121 |
+
elif want_next:
|
122 |
+
new_msgs.append(m)
|
123 |
+
want_next = False
|
124 |
+
error_msg = "\n".join(new_msgs)
|
125 |
+
return False, error_msg.strip()
|
126 |
+
|
127 |
+
def __call__(self, query: str) -> Tuple[bool, str]:
|
128 |
+
with ThreadPoolExecutor() as executor:
|
129 |
+
future = executor.submit(self.execute, query)
|
130 |
+
try:
|
131 |
+
return future.result(timeout=self.timeout)
|
132 |
+
except TimeoutError:
|
133 |
+
return False, f"Timed out after {self.timeout} seconds."
|
134 |
+
|
135 |
+
|
136 |
+
def execute_completion(
|
137 |
+
executor: PythonREPL,
|
138 |
+
completion: str,
|
139 |
+
return_status: bool = False,
|
140 |
+
last_code_block: bool = False,
|
141 |
+
) -> str | Tuple[str, bool]:
|
142 |
+
# executions = ["!" + code for code in re.findall(r"```bash(.*?)```", completion, re.DOTALL) if "!" not in code]
|
143 |
+
executions = re.findall(r"```python(.*?)```", completion, re.DOTALL)
|
144 |
+
|
145 |
+
if len(executions) == 0: # directly return cot result
|
146 |
+
return completion, False if return_status else completion
|
147 |
+
else:
|
148 |
+
if last_code_block:
|
149 |
+
executions = [executions[-1]]
|
150 |
+
|
151 |
+
# Python
|
152 |
+
execution_outputs = []
|
153 |
+
successes = []
|
154 |
+
for code in executions:
|
155 |
+
success = False
|
156 |
+
|
157 |
+
if "subprocess" in code:
|
158 |
+
output = "subprocess is not allowed"
|
159 |
+
execution_outputs.append(output)
|
160 |
+
successes.append(success)
|
161 |
+
continue
|
162 |
+
|
163 |
+
if "venv" in code:
|
164 |
+
output = "venv is not allowed"
|
165 |
+
execution_outputs.append(output)
|
166 |
+
successes.append(success)
|
167 |
+
continue
|
168 |
+
|
169 |
+
try:
|
170 |
+
success, output = executor(code)
|
171 |
+
except TimeoutError as e:
|
172 |
+
print("time out")
|
173 |
+
output = e
|
174 |
+
|
175 |
+
if not success and not return_status:
|
176 |
+
output = ""
|
177 |
+
|
178 |
+
execution_outputs.append(output)
|
179 |
+
successes.append(success)
|
180 |
+
|
181 |
+
output = str(execution_outputs[-1]).strip()
|
182 |
+
success = successes[-1]
|
183 |
+
|
184 |
+
if return_status:
|
185 |
+
return output, success
|
186 |
+
else:
|
187 |
+
return output
|
188 |
+
|
189 |
+
|
190 |
+
def postprocess_completion(
|
191 |
+
text: str, return_status: bool = False, last_code_block=False, timeout=5
|
192 |
+
) -> str | Tuple[str, bool]:
|
193 |
+
executor = PythonREPL(timeout=timeout)
|
194 |
+
|
195 |
+
result = execute_completion(executor, text, return_status=return_status, last_code_block=last_code_block)
|
196 |
+
del executor
|
197 |
+
|
198 |
+
return result
|
199 |
+
|
200 |
+
|
201 |
+
def apply_template(example: Dict[str, Any], prompt: str) -> Dict[str, Any]:
|
202 |
+
return prompt.format(example["prompt"], "{}")
|
203 |
+
|
204 |
+
|
205 |
+
def last_boxed_only_string(string):
|
206 |
+
"""
|
207 |
+
Extracts the last LaTeX boxed or framed expression from a string.
|
208 |
+
Args:
|
209 |
+
string (str): The input string containing LaTeX expressions.
|
210 |
+
Returns:
|
211 |
+
str or None: The last boxed or framed expression, if found;
|
212 |
+
otherwise, None.
|
213 |
+
"""
|
214 |
+
|
215 |
+
idx = string.rfind("\\boxed")
|
216 |
+
if idx < 0:
|
217 |
+
idx = string.rfind("\\fbox")
|
218 |
+
if idx < 0:
|
219 |
+
return None
|
220 |
+
|
221 |
+
i = idx
|
222 |
+
right_brace_idx = None
|
223 |
+
num_left_braces_open = 0
|
224 |
+
while i < len(string):
|
225 |
+
if string[i] == "{":
|
226 |
+
num_left_braces_open += 1
|
227 |
+
if string[i] == "}":
|
228 |
+
num_left_braces_open -= 1
|
229 |
+
if num_left_braces_open == 0:
|
230 |
+
right_brace_idx = i
|
231 |
+
break
|
232 |
+
i += 1
|
233 |
+
|
234 |
+
if right_brace_idx is None:
|
235 |
+
retval = None
|
236 |
+
else:
|
237 |
+
retval = string[idx : right_brace_idx + 1]
|
238 |
+
|
239 |
+
return retval
|
240 |
+
|
241 |
+
|
242 |
+
def remove_boxed(s):
|
243 |
+
"""
|
244 |
+
Removes the LaTeX boxed command, returning the content inside the braces.
|
245 |
+
Args:
|
246 |
+
s (str): The string containing a LaTeX boxed expression.
|
247 |
+
Returns:
|
248 |
+
str or None: The content inside the boxed command, if valid;
|
249 |
+
otherwise, None.
|
250 |
+
"""
|
251 |
+
|
252 |
+
left = "\\boxed{"
|
253 |
+
try:
|
254 |
+
assert s[: len(left)] == left
|
255 |
+
assert s[-1] == "}"
|
256 |
+
length = len(left)
|
257 |
+
return s[length:-1]
|
258 |
+
except Exception:
|
259 |
+
return None
|
260 |
+
|
261 |
+
|
262 |
+
def extract_boxed_answer(pred_str, strip_double_curly_brace=False):
|
263 |
+
"""
|
264 |
+
Extracts the answer from a LaTeX boxed expression within
|
265 |
+
a prediction string.
|
266 |
+
Args:
|
267 |
+
pred_str (str): The string containing one or more LaTeX
|
268 |
+
boxed expressions.
|
269 |
+
strip_double_curly_brace (bool): If True, removes an additional
|
270 |
+
layer of braces.
|
271 |
+
Returns:
|
272 |
+
str or None: The extracted answer, if any; otherwise, None.
|
273 |
+
"""
|
274 |
+
|
275 |
+
boxed_str = last_boxed_only_string(pred_str)
|
276 |
+
if boxed_str is None:
|
277 |
+
return None
|
278 |
+
answer = remove_boxed(boxed_str)
|
279 |
+
if answer is None:
|
280 |
+
return None
|
281 |
+
if strip_double_curly_brace:
|
282 |
+
match = re.match("^\{(.*)\}$", answer) # noqa: W605
|
283 |
+
if match:
|
284 |
+
answer = match.group(1)
|
285 |
+
return answer
|
286 |
+
|
287 |
+
|
288 |
+
def normalize_final_answer(final_answer: str) -> str:
|
289 |
+
"""
|
290 |
+
Normalizes a final answer string by removing or replacing various LaTeX
|
291 |
+
and text elements.
|
292 |
+
Args:
|
293 |
+
final_answer (str): The answer string to normalize.
|
294 |
+
Returns:
|
295 |
+
str: The normalized answer string.
|
296 |
+
"""
|
297 |
+
|
298 |
+
match = re.search(r"(.*?)Problem:", final_answer, flags=re.S)
|
299 |
+
if match:
|
300 |
+
final_answer = match.group(1) # 返回匹配的第一部分,即"Problem"之前的所有文本
|
301 |
+
"""Normalize a final answer to a quantitative reasoning question."""
|
302 |
+
# final_answer = final_answer.split('=')[-1]
|
303 |
+
SUBSTITUTIONS = [
|
304 |
+
("an ", ""),
|
305 |
+
("a ", ""),
|
306 |
+
(".$", "$"),
|
307 |
+
("\\$", ""),
|
308 |
+
(r"\ ", ""),
|
309 |
+
(" ", ""),
|
310 |
+
("mbox", "text"),
|
311 |
+
(",\\text{and}", ","),
|
312 |
+
("\\text{and}", ","),
|
313 |
+
("\\text{m}", "\\text{}"),
|
314 |
+
("\\le", "<"),
|
315 |
+
]
|
316 |
+
REMOVED_EXPRESSIONS = [
|
317 |
+
"square",
|
318 |
+
"ways",
|
319 |
+
"integers",
|
320 |
+
"dollars",
|
321 |
+
"mph",
|
322 |
+
"inches",
|
323 |
+
"ft",
|
324 |
+
"hours",
|
325 |
+
"km",
|
326 |
+
"units",
|
327 |
+
"\\ldots",
|
328 |
+
"sue",
|
329 |
+
"points",
|
330 |
+
"feet",
|
331 |
+
"minutes",
|
332 |
+
"digits",
|
333 |
+
"cents",
|
334 |
+
"degrees",
|
335 |
+
"cm",
|
336 |
+
"gm",
|
337 |
+
"pounds",
|
338 |
+
"meters",
|
339 |
+
"meals",
|
340 |
+
"edges",
|
341 |
+
"students",
|
342 |
+
"childrentickets",
|
343 |
+
"multiples",
|
344 |
+
"\\text{s}",
|
345 |
+
"\\text{.}",
|
346 |
+
"\\text{\ns}",
|
347 |
+
"\\text{}^2",
|
348 |
+
"\\text{}^3",
|
349 |
+
"\\text{\n}",
|
350 |
+
"\\text{}",
|
351 |
+
r"\mathrm{th}",
|
352 |
+
r"^\circ",
|
353 |
+
r"^{\circ}",
|
354 |
+
r"\;",
|
355 |
+
r",\!",
|
356 |
+
"{,}",
|
357 |
+
'"',
|
358 |
+
"\\dots",
|
359 |
+
"\n",
|
360 |
+
"\r",
|
361 |
+
"\f",
|
362 |
+
"\%",
|
363 |
+
]
|
364 |
+
for before, after in SUBSTITUTIONS:
|
365 |
+
final_answer = final_answer.replace(before, after)
|
366 |
+
for expr in REMOVED_EXPRESSIONS:
|
367 |
+
final_answer = final_answer.replace(expr, "")
|
368 |
+
|
369 |
+
# Extract answer that is in LaTeX math, is bold,
|
370 |
+
# is surrounded by a box, etc.
|
371 |
+
final_answer = re.sub(r"(\\text\{)(.*?)(\})", "\\2", final_answer)
|
372 |
+
final_answer = re.sub(r"(\\textbf\{)(.*?)(\})", "\\2", final_answer)
|
373 |
+
final_answer = re.sub(r"(\\overline\{)(.*?)(\})", "\\2", final_answer)
|
374 |
+
final_answer = re.sub(r"(\\boxed\{)(.*)(\})", "\\2", final_answer)
|
375 |
+
assert "\n" not in final_answer
|
376 |
+
assert "\r" not in final_answer
|
377 |
+
assert "\f" not in final_answer
|
378 |
+
if len(re.findall(r"finalansweris(.*)", final_answer)) > 0:
|
379 |
+
final_answer = re.findall(r"finalansweris(.*)", final_answer)[-1]
|
380 |
+
|
381 |
+
if len(re.findall(r"answer?is:?(.*)", final_answer)) > 0:
|
382 |
+
final_answer = re.findall(r"answer?is:?(.*)", final_answer)[-1]
|
383 |
+
|
384 |
+
if len(re.findall(r"oxed\{(.*?)\}", final_answer)) > 0:
|
385 |
+
final_answer = re.findall(r"oxed\{(.*?)\}", final_answer)[-1]
|
386 |
+
|
387 |
+
if len(re.findall(r"\$(.*?)\$", final_answer)) > 0:
|
388 |
+
final_answer = re.findall(r"\$(.*?)\$", final_answer)[-1]
|
389 |
+
final_answer = final_answer.strip()
|
390 |
+
if "rac" in final_answer and "\\frac" not in final_answer:
|
391 |
+
final_answer = final_answer.replace("rac", "\\frac")
|
392 |
+
|
393 |
+
final_answer = re.sub(r"(frac)([^{])(.)", "frac{\\2}{\\3}", final_answer)
|
394 |
+
final_answer = re.sub(r"(sqrt)([^{])", "sqrt{\\2}", final_answer)
|
395 |
+
final_answer = final_answer.replace("$", "")
|
396 |
+
|
397 |
+
if final_answer.replace(",", "").isdigit():
|
398 |
+
final_answer = final_answer.replace(",", "")
|
399 |
+
|
400 |
+
return final_answer
|
401 |
+
|
402 |
+
|
403 |
+
def naive_parse(answer: str) -> str:
|
404 |
+
"""
|
405 |
+
Extracts and returns the numeric digits from the input string, processing them in reverse order
|
406 |
+
until a non-numeric character is encountered after encountering the first numeric character.
|
407 |
+
|
408 |
+
Args:
|
409 |
+
answer (str): The input string to parse.
|
410 |
+
|
411 |
+
Returns:
|
412 |
+
str: A string consisting of the numeric digits extracted from the input, in their original order.
|
413 |
+
|
414 |
+
Example:
|
415 |
+
>>> naive_parse("abc123def")
|
416 |
+
'123'
|
417 |
+
>>> naive_parse("def456ghi")
|
418 |
+
'456'
|
419 |
+
>>> naive_parse("no numbers here")
|
420 |
+
''
|
421 |
+
"""
|
422 |
+
out = []
|
423 |
+
start = False
|
424 |
+
end = False
|
425 |
+
for l in reversed(list(answer)):
|
426 |
+
if l in "0123456789" and not end:
|
427 |
+
start = True
|
428 |
+
out.append(l)
|
429 |
+
else:
|
430 |
+
if start:
|
431 |
+
end = True
|
432 |
+
|
433 |
+
out = reversed(out)
|
434 |
+
return "".join(out)
|
435 |
+
|
436 |
+
|
437 |
+
def validate_answer_is_numeric(x: str | int | float) -> int:
|
438 |
+
FLOAT_TOLERANCE = 0.2
|
439 |
+
try:
|
440 |
+
x = round(float(x))
|
441 |
+
f = float(x)
|
442 |
+
if abs(x - f) > FLOAT_TOLERANCE:
|
443 |
+
x = -1
|
444 |
+
except Exception:
|
445 |
+
x = -1
|
446 |
+
return x
|
447 |
+
|
448 |
+
|
449 |
+
def get_majority_vote(responses: List[int]) -> int:
|
450 |
+
if len(responses) < 1:
|
451 |
+
return 0
|
452 |
+
else:
|
453 |
+
c = Counter(responses)
|
454 |
+
value, count = c.most_common()[0]
|
455 |
+
return value
|
456 |
+
|
457 |
+
|
458 |
+
def filter_answers(answers: List[str]) -> List[int]:
|
459 |
+
formatted_answers = [validate_answer_is_numeric(a) for a in answers]
|
460 |
+
|
461 |
+
# Filter for non-negative answers
|
462 |
+
formatted_answers = [a for a in formatted_answers if a >= 0]
|
463 |
+
# Compute modulo
|
464 |
+
formatted_answers = [a % 1_000 for a in formatted_answers]
|
465 |
+
# less than 2.1 billion or cannot convert to C int (32-bit)
|
466 |
+
formatted_answers = [a for a in formatted_answers if a <= 999]
|
467 |
+
return formatted_answers
|
468 |
+
|
469 |
+
|
470 |
+
def check_sympy_equivalence(ref_answer: str, model_answer: str) -> bool:
|
471 |
+
def do_answers_match(ref_answer: str, model_answer: str) -> bool:
|
472 |
+
ref_sympy = parse_latex(ref_answer)
|
473 |
+
model_sympy = parse_latex(model_answer)
|
474 |
+
diff = simplify(ref_sympy - model_sympy)
|
475 |
+
return True if -1e-12 < N(diff) < 1e-12 or diff.is_zero else False
|
476 |
+
|
477 |
+
try:
|
478 |
+
result = do_answers_match(ref_answer, model_answer)
|
479 |
+
return result
|
480 |
+
except Exception as e:
|
481 |
+
print(e)
|
482 |
+
return False
|
483 |
+
|
484 |
+
|
485 |
+
def check_string_match(ref_answer: str, model_answer: str) -> bool:
|
486 |
+
try:
|
487 |
+
return ref_answer == model_answer
|
488 |
+
except Exception as e:
|
489 |
+
print(e)
|
490 |
+
return False
|
491 |
+
|
492 |
+
|
493 |
+
def check_answer(ref_answer: str, model_answer: str) -> bool:
|
494 |
+
# check if strings are the same
|
495 |
+
correct = check_string_match(ref_answer, model_answer)
|
496 |
+
if correct:
|
497 |
+
return True
|
498 |
+
|
499 |
+
# use the sympy library to check if the expressions are the same
|
500 |
+
correct = check_sympy_equivalence(ref_answer, model_answer)
|
501 |
+
if correct:
|
502 |
+
return True
|
503 |
+
|
504 |
+
return False
|
505 |
+
|
506 |
+
|
507 |
+
debug = False
|
508 |
+
model_id = "Numina-Math-7B"
|
509 |
+
revision = "main"
|
510 |
+
system_prompt = "{}"
|
511 |
+
validation_set = "kaggle-validation-set-medium"
|
512 |
+
is_submission = True
|
513 |
+
num_samples = 4
|
514 |
+
num_generations = 4
|
515 |
+
temperature = 0.8
|
516 |
+
is_quantized = False
|
517 |
+
restart_on_fail = False
|
518 |
+
top_p = 1.0
|
519 |
+
top_k = 0
|
520 |
+
max_new_tokens = 2048
|
521 |
+
# Papermill related variables
|
522 |
+
push_to_hub = False
|
523 |
+
notebook_name = ""
|
524 |
+
|
525 |
+
config = Config(
|
526 |
+
debug=debug,
|
527 |
+
push_to_hub=push_to_hub,
|
528 |
+
model_id=model_id,
|
529 |
+
revision=revision,
|
530 |
+
system_prompt=system_prompt,
|
531 |
+
validation_set=validation_set,
|
532 |
+
is_quantized=is_quantized,
|
533 |
+
restart_on_fail=restart_on_fail,
|
534 |
+
is_submission=is_submission,
|
535 |
+
num_samples=num_samples,
|
536 |
+
num_generations=num_generations,
|
537 |
+
do_sample=True,
|
538 |
+
temperature=temperature,
|
539 |
+
top_p=top_p,
|
540 |
+
top_k=top_k,
|
541 |
+
max_new_tokens=max_new_tokens,
|
542 |
+
)
|
543 |
+
print(f"=== Running submission with config ===\n\n{config}")
|
544 |
+
|
545 |
+
|
546 |
+
def generate(message):
|
547 |
+
chat_completion = client.chat.completions.create(
|
548 |
+
model="tgi",
|
549 |
+
messages=message,
|
550 |
+
stream=True,
|
551 |
+
max_tokens=1024,
|
552 |
+
stop=["```output\n"],
|
553 |
+
temperature=temperature,
|
554 |
+
)
|
555 |
+
|
556 |
+
for message in chat_completion:
|
557 |
+
yield message.choices[0].delta.content
|
558 |
+
|
559 |
+
|
560 |
+
def get_majority_text(data):
|
561 |
+
from collections import Counter
|
562 |
+
|
563 |
+
# Count the frequency of each answer in model_answers
|
564 |
+
answer_counts = Counter(data["model_answers"])
|
565 |
+
|
566 |
+
# Find the majority response
|
567 |
+
majority_response = answer_counts.most_common(1)[0][0]
|
568 |
+
|
569 |
+
# Find the index of the first occurrence of the majority response
|
570 |
+
majority_index = data["model_answers"].index(majority_response)
|
571 |
+
|
572 |
+
# Return the corresponding text in gen_texts
|
573 |
+
return data["gen_texts"][majority_index]
|
574 |
+
|
575 |
+
|
576 |
+
def extract_solution(text):
|
577 |
+
# Split the text at "### Solution:"
|
578 |
+
parts = text.split("### Solution:", 1)
|
579 |
+
if len(parts) > 1:
|
580 |
+
# Return everything after "### Solution:"
|
581 |
+
return parts[1].strip()
|
582 |
+
else:
|
583 |
+
# Return an empty string if "### Solution:" is not found
|
584 |
+
return ""
|
585 |
+
|
586 |
+
|
587 |
+
def process_code(
|
588 |
+
example: Dict[str, Any],
|
589 |
+
config: Config,
|
590 |
+
restart_on_fail: bool = False,
|
591 |
+
last_step: bool = False,
|
592 |
+
) -> Dict[str, Any]:
|
593 |
+
gen_text = example["gen_texts"]
|
594 |
+
num_python_blocks = len(re.findall(r"```python(.*?)```", gen_text, re.DOTALL))
|
595 |
+
|
596 |
+
if num_python_blocks == 0:
|
597 |
+
if restart_on_fail:
|
598 |
+
print("no code has ever been generated, RESTARTING")
|
599 |
+
# reset the text to the original
|
600 |
+
example["gen_texts"] = example["text"]
|
601 |
+
else:
|
602 |
+
print("no code has ever been generated, STOP")
|
603 |
+
example["should_prune"] = True
|
604 |
+
example["has_code"] = False
|
605 |
+
return example
|
606 |
+
|
607 |
+
if gen_text[-10:] != "```output\n" and ("answer is" in gen_text[-100:] or "\\boxed" in gen_text[-100:]):
|
608 |
+
num_output_blocks = len(re.findall(r"```output(.*?)```", gen_text, re.DOTALL))
|
609 |
+
if num_output_blocks == 0:
|
610 |
+
print("the model hallucinated the code answer")
|
611 |
+
example["should_prune"] = True
|
612 |
+
return example
|
613 |
+
|
614 |
+
if "boxed" in gen_text[-100:]:
|
615 |
+
try:
|
616 |
+
answer = normalize_final_answer(extract_boxed_answer(gen_text[-100:]))
|
617 |
+
except Exception:
|
618 |
+
answer = "-1"
|
619 |
+
else:
|
620 |
+
answer = normalize_final_answer(gen_text[-100:])
|
621 |
+
|
622 |
+
example["model_answers"] = answer
|
623 |
+
if not config.is_submission:
|
624 |
+
example["corrects"] = check_answer(example["ground_truth"], answer)
|
625 |
+
example["should_prune"] = True
|
626 |
+
print("Answer is: ", answer, example["ground_truth"], example["corrects"])
|
627 |
+
return example
|
628 |
+
|
629 |
+
if last_step:
|
630 |
+
# no point in continuing if we are at the last step
|
631 |
+
return example
|
632 |
+
|
633 |
+
if gen_text[-10:] != "```output\n":
|
634 |
+
# something else has gone wrong with the generation
|
635 |
+
print("warning: output block not found: ", gen_text[-40:])
|
636 |
+
if restart_on_fail:
|
637 |
+
example["gen_texts"] = example["text"]
|
638 |
+
else:
|
639 |
+
example["should_prune"] = True
|
640 |
+
return example
|
641 |
+
|
642 |
+
code_result, status = postprocess_completion(gen_text, return_status=True, last_code_block=True)
|
643 |
+
# add the code result for the next round of generation
|
644 |
+
TRUNCATION_LIMIT = 200
|
645 |
+
if len(code_result) > TRUNCATION_LIMIT:
|
646 |
+
code_result = code_result[:TRUNCATION_LIMIT] + " ... (output truncated)"
|
647 |
+
example["gen_texts"] = gen_text + f"{code_result}\n```"
|
648 |
+
|
649 |
+
return example
|
650 |
+
|
651 |
+
|
652 |
+
# load the vllm instance and set sampling parameters
|
653 |
+
# vllm = build_vllm(config)
|
654 |
+
|
655 |
+
|
656 |
+
def solve_problem(problem, temperature, progress=gr.Progress()):
|
657 |
+
problem = apply_template({"prompt": problem}, prompt=config.system_prompt)
|
658 |
+
print(f"Problem: {problem}")
|
659 |
+
|
660 |
+
sample = {
|
661 |
+
"problem": problem, # not used for the submission TODO Remove
|
662 |
+
"ground_truth": "unknown", # not used for the submission TODO Remove
|
663 |
+
"text": "### Solution:\n",
|
664 |
+
"gen_texts": "### Solution:\n", # used to store all the generated text
|
665 |
+
"should_prune": False,
|
666 |
+
"problem_index": -1, # not used for the submission TODO Remove
|
667 |
+
"model_answers": "-1",
|
668 |
+
"has_code": True,
|
669 |
+
"corrects": False, # not used for the submission TODO Remove
|
670 |
+
}
|
671 |
+
|
672 |
+
for step in progress.tqdm(
|
673 |
+
range(config.num_generations), desc="Generating candidates"
|
674 |
+
): # Depth of the tree (e.g. 6 steps = 5 code blocks)
|
675 |
+
|
676 |
+
step_reponse = sample["gen_texts"]
|
677 |
+
|
678 |
+
messages = [
|
679 |
+
{"role": "user", "content": sample["problem"]},
|
680 |
+
{"role": "assistant", "content": sample["gen_texts"]},
|
681 |
+
]
|
682 |
+
|
683 |
+
for reponse_message in generate(messages, temperature):
|
684 |
+
if reponse_message is not None:
|
685 |
+
step_reponse += reponse_message
|
686 |
+
yield step_reponse
|
687 |
+
|
688 |
+
sample["gen_texts"] = step_reponse
|
689 |
+
|
690 |
+
# TODO: Maybe it should just return the result of running the code
|
691 |
+
sample = process_code(
|
692 |
+
sample,
|
693 |
+
config=config,
|
694 |
+
restart_on_fail=config.restart_on_fail,
|
695 |
+
last_step=(step == (config.num_generations - 1)),
|
696 |
+
)
|
697 |
+
sample["gen_texts"] = sample["gen_texts"] + "\n"
|
698 |
+
|
699 |
+
run_code_reponse = sample["gen_texts"].replace(step_reponse, "")
|
700 |
+
|
701 |
+
for output_mseeage in run_code_reponse:
|
702 |
+
if output_mseeage is not None:
|
703 |
+
step_reponse += output_mseeage
|
704 |
+
yield step_reponse
|
705 |
+
|
706 |
+
if sample["should_prune"]:
|
707 |
+
break
|
708 |
+
|
709 |
+
yield sample["gen_texts"]
|
710 |
+
|
711 |
+
|
712 |
+
with gr.Blocks() as demo:
|
713 |
+
with gr.Row():
|
714 |
+
inp = gr.Textbox(placeholder="Problem", label="Problem", lines=5)
|
715 |
+
with gr.Accordion("Advanced Options", open=False):
|
716 |
+
temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.8, step=0.1, label="Temperature")
|
717 |
+
with gr.Row():
|
718 |
+
out = gr.Markdown()
|
719 |
+
|
720 |
+
btn = gr.Button("Run")
|
721 |
+
btn.click(fn=solve_problem, inputs=[inp, temperature], outputs=out)
|
722 |
+
|
723 |
+
|
724 |
+
if __name__ == "__main__":
|
725 |
+
demo.queue(default_concurrency_limit=5).launch()
|