feat: add math parser
Browse files- .pre-commit-config.yaml +0 -4
- llmdataparser/math_parser.py +108 -0
- pyproject.toml +0 -3
- tests/test_math_parser.py +200 -0
.pre-commit-config.yaml
CHANGED
@@ -65,14 +65,10 @@ repos:
|
|
65 |
- id: prettier
|
66 |
types_or: [markdown, yaml]
|
67 |
- repo: https://github.com/astral-sh/ruff-pre-commit
|
68 |
-
# Ruff version.
|
69 |
rev: v0.4.4
|
70 |
hooks:
|
71 |
-
# Run the linter.
|
72 |
- id: ruff
|
73 |
args: [--fix]
|
74 |
-
# Run the formatter.
|
75 |
-
- id: ruff-format
|
76 |
- repo: https://github.com/kynan/nbstripout
|
77 |
rev: 0.5.0 # use the latest version
|
78 |
hooks:
|
|
|
65 |
- id: prettier
|
66 |
types_or: [markdown, yaml]
|
67 |
- repo: https://github.com/astral-sh/ruff-pre-commit
|
|
|
68 |
rev: v0.4.4
|
69 |
hooks:
|
|
|
70 |
- id: ruff
|
71 |
args: [--fix]
|
|
|
|
|
72 |
- repo: https://github.com/kynan/nbstripout
|
73 |
rev: 0.5.0 # use the latest version
|
74 |
hooks:
|
llmdataparser/math_parser.py
ADDED
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from typing import Any, ClassVar
|
3 |
+
|
4 |
+
from llmdataparser.base_parser import HuggingFaceDatasetParser, HuggingFaceParseEntry
|
5 |
+
|
6 |
+
|
7 |
+
@dataclass(frozen=True, kw_only=True, slots=True)
|
8 |
+
class MATHParseEntry(HuggingFaceParseEntry):
|
9 |
+
"""Custom entry class for MATH dataset, with fields specific to this dataset parser."""
|
10 |
+
|
11 |
+
level: str
|
12 |
+
task_name: str
|
13 |
+
solution: str
|
14 |
+
|
15 |
+
@classmethod
|
16 |
+
def create(
|
17 |
+
cls,
|
18 |
+
prompt: str,
|
19 |
+
answer: str,
|
20 |
+
raw_question: str,
|
21 |
+
raw_answer: str,
|
22 |
+
level: str,
|
23 |
+
task_name: str,
|
24 |
+
solution: str,
|
25 |
+
) -> "MATHParseEntry":
|
26 |
+
return cls(
|
27 |
+
prompt=prompt,
|
28 |
+
answer=answer,
|
29 |
+
raw_question=raw_question,
|
30 |
+
raw_answer=raw_answer,
|
31 |
+
level=level,
|
32 |
+
task_name=task_name,
|
33 |
+
solution=solution,
|
34 |
+
)
|
35 |
+
|
36 |
+
|
37 |
+
class MATHDatasetParser(HuggingFaceDatasetParser[MATHParseEntry]):
|
38 |
+
"""Parser for the MATH dataset."""
|
39 |
+
|
40 |
+
_data_source: ClassVar[str] = "lighteval/MATH"
|
41 |
+
_task_names: ClassVar[list[str]] = [
|
42 |
+
"algebra",
|
43 |
+
"geometry",
|
44 |
+
"calculus",
|
45 |
+
"prealgebra",
|
46 |
+
"intermediate_algebra",
|
47 |
+
"number_theory",
|
48 |
+
"precalculus",
|
49 |
+
"all",
|
50 |
+
]
|
51 |
+
_default_task: ClassVar[str] = "all"
|
52 |
+
_default_system_prompt: ClassVar[
|
53 |
+
str
|
54 |
+
] = "Solve the following mathematics problem step by step:"
|
55 |
+
_valid_levels: ClassVar[set[str]] = {
|
56 |
+
f"Level {i}" for i in range(1, 6)
|
57 |
+
} # Levels 1-5 are valid
|
58 |
+
|
59 |
+
def _get_task_from_entry(self, data_entry: dict[str, Any]) -> str:
|
60 |
+
"""Get the task name from the data entry or fall back to current task."""
|
61 |
+
entry_type = data_entry.get("type")
|
62 |
+
if entry_type and (entry_type in self._task_names):
|
63 |
+
return entry_type
|
64 |
+
return self._current_task or self._default_task
|
65 |
+
|
66 |
+
def process_entry(
|
67 |
+
self, row: dict[str, Any], task_name: str | None = None, **kwargs: Any
|
68 |
+
) -> MATHParseEntry:
|
69 |
+
"""Process a single MATH dataset entry."""
|
70 |
+
task = task_name or self._get_current_task(row)
|
71 |
+
|
72 |
+
# Validate and normalize level
|
73 |
+
level = row.get("level")
|
74 |
+
if level not in self._valid_levels:
|
75 |
+
level = "Unknown"
|
76 |
+
|
77 |
+
return MATHParseEntry.create(
|
78 |
+
prompt=f"{self._system_prompt}\n{row['problem']}",
|
79 |
+
answer=row["solution"],
|
80 |
+
raw_question=row["problem"],
|
81 |
+
raw_answer=row["solution"],
|
82 |
+
level=level,
|
83 |
+
task_name=task,
|
84 |
+
solution=row["solution"],
|
85 |
+
)
|
86 |
+
|
87 |
+
|
88 |
+
if __name__ == "__main__":
|
89 |
+
# Example usage of MATH parser
|
90 |
+
parser = MATHDatasetParser()
|
91 |
+
|
92 |
+
# Load the dataset
|
93 |
+
parser.load()
|
94 |
+
|
95 |
+
# Parse all splits
|
96 |
+
parser.parse()
|
97 |
+
|
98 |
+
# Get parsed data
|
99 |
+
parsed_data = parser.get_parsed_data
|
100 |
+
|
101 |
+
# Print example entry
|
102 |
+
if parsed_data:
|
103 |
+
example = parsed_data[0]
|
104 |
+
print("\nExample parsed entry:")
|
105 |
+
print(f"Task: {example.task_name}")
|
106 |
+
print(f"Level: {example.level}")
|
107 |
+
print(f"Question: {example.raw_question}")
|
108 |
+
print(f"Solution: {example.solution}")
|
pyproject.toml
CHANGED
@@ -49,11 +49,8 @@ profile = "black"
|
|
49 |
line_length = 88
|
50 |
known_first_party = ["llmdataparser"]
|
51 |
|
52 |
-
|
53 |
[tool.ruff]
|
54 |
line-length = 88
|
55 |
-
select = ["E", "F"] # or specify checks explicitly without E501
|
56 |
-
ignore = ["E501"]
|
57 |
|
58 |
[tool.ruff.lint]
|
59 |
ignore = ["E501"]
|
|
|
49 |
line_length = 88
|
50 |
known_first_party = ["llmdataparser"]
|
51 |
|
|
|
52 |
[tool.ruff]
|
53 |
line-length = 88
|
|
|
|
|
54 |
|
55 |
[tool.ruff.lint]
|
56 |
ignore = ["E501"]
|
tests/test_math_parser.py
ADDED
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pytest
|
2 |
+
|
3 |
+
from llmdataparser.math_parser import MATHDatasetParser, MATHParseEntry
|
4 |
+
|
5 |
+
|
6 |
+
@pytest.fixture
|
7 |
+
def math_parser():
|
8 |
+
"""Create a MATH parser instance for testing."""
|
9 |
+
return MATHDatasetParser()
|
10 |
+
|
11 |
+
|
12 |
+
@pytest.fixture
|
13 |
+
def loaded_math_parser(math_parser):
|
14 |
+
"""Create and load a MATH parser instance with test split."""
|
15 |
+
math_parser.load(task_name="algebra", split="test")
|
16 |
+
return math_parser
|
17 |
+
|
18 |
+
|
19 |
+
@pytest.fixture
|
20 |
+
def sample_math_entries():
|
21 |
+
"""Create sample MATH dataset entries for testing."""
|
22 |
+
return [
|
23 |
+
{
|
24 |
+
"problem": "Solve for x: 2x + 4 = 10",
|
25 |
+
"level": "Level 3",
|
26 |
+
"solution": "Let's solve step by step:\n1) Subtract 4 from both sides: 2x = 6\n2) Divide both sides by 2\n\nTherefore, x = 3",
|
27 |
+
"type": "algebra",
|
28 |
+
},
|
29 |
+
{
|
30 |
+
"problem": "Find the area of a circle with radius 5 units.",
|
31 |
+
"level": "Level 2",
|
32 |
+
"solution": "Area = πr²\nArea = π(5)²\nArea = 25π square units",
|
33 |
+
"type": "geometry",
|
34 |
+
},
|
35 |
+
{
|
36 |
+
"problem": "What is the limit of (x²-1)/(x-1) as x approaches 1?",
|
37 |
+
"level": "Level 4",
|
38 |
+
"solution": "Using L'Hôpital's rule:\nlim(x→1) (x²-1)/(x-1) = lim(x→1) (2x)/(1) = 2",
|
39 |
+
"type": "calculus",
|
40 |
+
},
|
41 |
+
]
|
42 |
+
|
43 |
+
|
44 |
+
def test_math_parse_entry_creation_valid():
|
45 |
+
"""Test valid creation of MATHParseEntry with all fields."""
|
46 |
+
entry = MATHParseEntry.create(
|
47 |
+
prompt="Test prompt",
|
48 |
+
answer="Test answer",
|
49 |
+
raw_question="Test question",
|
50 |
+
raw_answer="Test solution",
|
51 |
+
level="Level 5",
|
52 |
+
task_name="algebra",
|
53 |
+
solution="Test solution",
|
54 |
+
)
|
55 |
+
|
56 |
+
assert isinstance(entry, MATHParseEntry)
|
57 |
+
assert entry.prompt == "Test prompt"
|
58 |
+
assert entry.answer == "Test answer"
|
59 |
+
assert entry.raw_question == "Test question"
|
60 |
+
assert entry.raw_answer == "Test solution"
|
61 |
+
assert entry.level == "Level 5"
|
62 |
+
assert entry.task_name == "algebra"
|
63 |
+
assert entry.solution == "Test solution"
|
64 |
+
|
65 |
+
|
66 |
+
@pytest.mark.parametrize(
|
67 |
+
"test_case",
|
68 |
+
[
|
69 |
+
{
|
70 |
+
"problem": "Solve for x: 2x + 4 = 10",
|
71 |
+
"level": "Level 3",
|
72 |
+
"solution": "x = 3",
|
73 |
+
"type": "algebra",
|
74 |
+
},
|
75 |
+
{
|
76 |
+
"problem": "Find the derivative of f(x) = x²",
|
77 |
+
"level": "Level 4",
|
78 |
+
"solution": "f'(x) = 2x",
|
79 |
+
"type": "calculus",
|
80 |
+
},
|
81 |
+
],
|
82 |
+
)
|
83 |
+
def test_process_entry(math_parser, test_case):
|
84 |
+
"""Test processing different types of MATH entries."""
|
85 |
+
entry = math_parser.process_entry(test_case, task_name=test_case["type"])
|
86 |
+
|
87 |
+
assert isinstance(entry, MATHParseEntry)
|
88 |
+
assert (
|
89 |
+
entry.prompt == f"{math_parser._default_system_prompt}\n{test_case['problem']}"
|
90 |
+
)
|
91 |
+
assert entry.answer == test_case["solution"]
|
92 |
+
assert entry.raw_question == test_case["problem"]
|
93 |
+
assert entry.raw_answer == test_case["solution"]
|
94 |
+
assert entry.level == test_case["level"]
|
95 |
+
assert entry.task_name == test_case["type"]
|
96 |
+
assert entry.solution == test_case["solution"]
|
97 |
+
|
98 |
+
|
99 |
+
def test_math_parser_initialization(math_parser):
|
100 |
+
"""Test MATH parser initialization and properties."""
|
101 |
+
assert isinstance(math_parser.task_names, list)
|
102 |
+
assert len(math_parser.task_names) == 8
|
103 |
+
assert math_parser._data_source == "lighteval/MATH"
|
104 |
+
assert math_parser._default_task == "all"
|
105 |
+
assert "algebra" in math_parser.task_names
|
106 |
+
assert "geometry" in math_parser.task_names
|
107 |
+
assert (
|
108 |
+
math_parser.get_huggingface_link
|
109 |
+
== "https://huggingface.co/datasets/lighteval/MATH"
|
110 |
+
)
|
111 |
+
assert "mathematics problem" in math_parser._default_system_prompt.lower()
|
112 |
+
|
113 |
+
|
114 |
+
def test_get_current_task(math_parser):
|
115 |
+
"""Test task name resolution in different scenarios."""
|
116 |
+
# Test with valid type in data entry
|
117 |
+
test_row_with_type = {"type": "algebra"}
|
118 |
+
assert math_parser._get_current_task(test_row_with_type) == "algebra"
|
119 |
+
|
120 |
+
# Test without type in data entry
|
121 |
+
test_row_without_type = {}
|
122 |
+
math_parser._current_task = "geometry"
|
123 |
+
assert math_parser._get_current_task(test_row_without_type) == "geometry"
|
124 |
+
|
125 |
+
# Test with invalid type - should return current task
|
126 |
+
test_row_invalid_type = {"type": "invalid_type"}
|
127 |
+
math_parser._current_task = "algebra"
|
128 |
+
assert math_parser._get_current_task(test_row_invalid_type) == "algebra"
|
129 |
+
|
130 |
+
|
131 |
+
def test_valid_levels(math_parser):
|
132 |
+
"""Test handling of valid level values."""
|
133 |
+
for i in range(1, 6):
|
134 |
+
test_row = {
|
135 |
+
"problem": "Test problem",
|
136 |
+
"level": f"Level {i}",
|
137 |
+
"solution": "Test solution",
|
138 |
+
"type": "algebra",
|
139 |
+
}
|
140 |
+
entry = math_parser.process_entry(test_row, task_name="algebra")
|
141 |
+
assert entry.level == f"Level {i}"
|
142 |
+
|
143 |
+
|
144 |
+
@pytest.mark.parametrize(
|
145 |
+
"invalid_level",
|
146 |
+
[
|
147 |
+
"Level 0", # Too low
|
148 |
+
"Level 6", # Too high
|
149 |
+
"Invalid", # Wrong format
|
150 |
+
None, # Missing
|
151 |
+
"", # Empty
|
152 |
+
"level 1", # Wrong capitalization
|
153 |
+
],
|
154 |
+
)
|
155 |
+
def test_invalid_level_handling(math_parser, invalid_level):
|
156 |
+
"""Test handling of invalid level values."""
|
157 |
+
test_row = {
|
158 |
+
"problem": "Test problem",
|
159 |
+
"level": invalid_level,
|
160 |
+
"solution": "Test solution",
|
161 |
+
"type": "algebra",
|
162 |
+
}
|
163 |
+
|
164 |
+
entry = math_parser.process_entry(test_row, task_name="algebra")
|
165 |
+
assert entry.level == "Unknown"
|
166 |
+
|
167 |
+
|
168 |
+
@pytest.mark.integration
|
169 |
+
def test_load_dataset(loaded_math_parser):
|
170 |
+
"""Test loading the MATH dataset."""
|
171 |
+
assert loaded_math_parser.raw_data is not None
|
172 |
+
assert loaded_math_parser.split_names == ["test"]
|
173 |
+
assert loaded_math_parser._current_task == "algebra"
|
174 |
+
|
175 |
+
|
176 |
+
def test_parser_string_representation(loaded_math_parser):
|
177 |
+
"""Test string representation of MATH parser."""
|
178 |
+
repr_str = str(loaded_math_parser)
|
179 |
+
assert "MATHDatasetParser" in repr_str
|
180 |
+
assert "lighteval/MATH" in repr_str
|
181 |
+
assert "algebra" in repr_str
|
182 |
+
assert "loaded" in repr_str
|
183 |
+
|
184 |
+
|
185 |
+
@pytest.mark.integration
|
186 |
+
def test_different_splits_parsing(math_parser):
|
187 |
+
"""Test parsing different splits of the dataset."""
|
188 |
+
# Load and parse test split
|
189 |
+
math_parser.load(task_name="algebra", split="test")
|
190 |
+
math_parser.parse(split_names="test", force=True)
|
191 |
+
test_count = len(math_parser.get_parsed_data)
|
192 |
+
|
193 |
+
# Load and parse train split
|
194 |
+
math_parser.load(task_name="algebra", split="train")
|
195 |
+
math_parser.parse(split_names="train", force=True)
|
196 |
+
train_count = len(math_parser.get_parsed_data)
|
197 |
+
|
198 |
+
assert test_count > 0
|
199 |
+
assert train_count > 0
|
200 |
+
assert train_count != test_count
|