Spaces:
Sleeping
Sleeping
File size: 14,812 Bytes
5db30ac | 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 | from __future__ import annotations
import json
import re
from typing import Any, Iterator
from elbench.loaders.base import BaseLoader
from elbench.registry import ResolvedRegistryItem
from elbench.schemas.evaluation import RawRecord, Sample
class MMLUProJsonlLoader(BaseLoader):
def iter_raw_records(self, item: ResolvedRegistryItem) -> Iterator[RawRecord]:
with item.path.open("r", encoding="utf-8") as handle:
for row_index, line in enumerate(handle, start=1):
if not line.strip():
continue
yield RawRecord(
source_path=str(item.path),
source_file=item.entry.canonical_name,
row_index=row_index,
record=json.loads(line),
)
def iter_samples(self, item: ResolvedRegistryItem) -> Iterator[Sample]:
for raw_record in self.iter_raw_records(item):
record = raw_record.record
target = self._string_or_none(record.get("target"))
choices = self._choices(record.get("choices"))
metadata = self._metadata(record, raw_record.row_index)
dimension = self._dimension(record, metadata)
yield Sample(
sample_id=self._sample_id(
item.entry.canonical_name,
record,
raw_record.row_index,
namespace=dimension,
),
source_file=item.entry.canonical_name,
source_path=str(item.path),
module=item.entry.module,
subset=item.entry.subset,
task=item.entry.task,
dimension=dimension,
prompt=self._build_zero_shot_prompt(record, choices, item.entry.task),
reference={"target": target},
metadata=metadata,
raw_record=record,
)
def _build_zero_shot_prompt(self, record: dict[str, Any], choices: list[str], task: str | None) -> str:
question = self._extract_question(record)
option_lines = [f"{chr(65 + index)}) {choice}" for index, choice in enumerate(choices)]
letters = ",".join(chr(65 + index) for index in range(len(choices)))
if task == "ceval":
return "\n".join(
[
"请回答以下单项选择题。",
"请先逐步思考，再给出答案。",
f"你回答的最后一行必须严格使用格式：答案：[LETTER]，其中 [LETTER] 是 {letters} 中的一个。",
"",
"问题：",
question,
"选项：",
*option_lines,
]
)
return "\n".join(
[
"Answer the following multiple choice question.",
"Think step by step before answering.",
f"The last line of your response must be exactly: ANSWER: [LETTER], where [LETTER] is one of {letters}.",
"",
"Question:",
question,
"Options:",
*option_lines,
]
)
def _extract_question(self, record: dict[str, Any]) -> str:
input_messages = record.get("input")
if isinstance(input_messages, list):
for message in reversed(input_messages):
if not isinstance(message, dict):
continue
content = message.get("content")
if not isinstance(content, str):
continue
matches = re.findall(r"Question:\n(.*?)\nOptions:", content, flags=re.S)
if not matches:
matches = re.findall(r"问题[：:]\s*(.*?)\n选项[：:]", content, flags=re.S)
if matches:
return matches[-1].strip()
raise ValueError("Cannot extract multiple-choice question from input messages.")
def _metadata(self, record: dict[str, Any], row_index: int) -> dict[str, Any]:
metadata = {}
raw_metadata = record.get("metadata")
if isinstance(raw_metadata, dict):
metadata.update(raw_metadata)
metadata["choices"] = self._choices(record.get("choices"))
metadata["subset_key"] = self._string_or_none(record.get("subset_key"))
metadata["row_index"] = row_index
return metadata
def _dimension(self, record: dict[str, Any], metadata: dict[str, Any]) -> str | None:
return (
self._string_or_none(record.get("subset_key"))
or self._string_or_none(metadata.get("subject"))
or self._string_or_none(metadata.get("id"))
)
def _sample_id(
self,
canonical_name: str,
record: dict[str, Any],
row_index: int,
namespace: str | None = None,
) -> str:
value = record.get("id")
if value in (None, ""):
value = row_index
value_part = self._slug_component(value)
namespace_part = self._slug_component(namespace)
if namespace_part:
return f"{canonical_name.replace('.', '_')}-{namespace_part}-{value_part}"
return f"{canonical_name.replace('.', '_')}-{value_part}"
def _choices(self, value: Any) -> list[str]:
if not isinstance(value, list) or not value:
raise ValueError("Multiple-choice record must contain a non-empty choices list.")
return [str(choice) for choice in value]
def _string_or_none(self, value: Any) -> str | None:
if value in (None, ""):
return None
return str(value)
def _slug_component(self, value: Any) -> str:
if value in (None, ""):
return ""
text = re.sub(r"[^\w]+", "_", str(value)).strip("_")
return text or str(value)
def expected_answer_from_sample(self, sample: Sample) -> str | None:
reference = sample.reference or {}
if isinstance(reference, dict):
value = reference.get("target")
else:
value = reference
if value in (None, ""):
return None
normalized = str(value).strip().upper()
return normalized[0] if normalized else None
def valid_letters_from_sample(self, sample: Sample) -> list[str]:
choices = sample.metadata.get("choices")
if isinstance(choices, list) and choices:
return [chr(65 + index) for index in range(len(choices))]
return list("ABCDEFGHIJ")
class CEvalJsonlLoader(MMLUProJsonlLoader):
pass
class IFEvalJsonlLoader(BaseLoader):
def iter_raw_records(self, item: ResolvedRegistryItem) -> Iterator[RawRecord]:
with item.path.open("r", encoding="utf-8") as handle:
for row_index, line in enumerate(handle, start=1):
if not line.strip():
continue
yield RawRecord(
source_path=str(item.path),
source_file=item.entry.canonical_name,
row_index=row_index,
record=json.loads(line),
)
def iter_samples(self, item: ResolvedRegistryItem) -> Iterator[Sample]:
for raw_record in self.iter_raw_records(item):
record = raw_record.record
metadata = self._metadata(record, raw_record.row_index)
yield Sample(
sample_id=self._sample_id(item.entry.canonical_name, record, raw_record.row_index),
source_file=item.entry.canonical_name,
source_path=str(item.path),
module=item.entry.module,
subset=item.entry.subset,
task=item.entry.task,
dimension="default",
prompt=self._prompt(record, metadata),
reference={"target": self._string_or_empty(record.get("target"))},
metadata=metadata,
raw_record=record,
)
def _metadata(self, record: dict[str, Any], row_index: int) -> dict[str, Any]:
metadata: dict[str, Any] = {}
raw_metadata = record.get("metadata")
if isinstance(raw_metadata, dict):
metadata.update(raw_metadata)
metadata["instruction_id_list"] = self._instruction_ids(metadata.get("instruction_id_list"))
metadata["kwargs"] = self._kwargs(metadata.get("kwargs"))
metadata["key"] = self._string_or_empty(metadata.get("key"))
metadata["prompt"] = self._prompt(record, metadata)
metadata["row_index"] = row_index
return metadata
def _prompt(self, record: dict[str, Any], metadata: dict[str, Any]) -> str:
prompt = metadata.get("prompt")
if isinstance(prompt, str) and prompt:
return prompt
input_messages = record.get("input")
if isinstance(input_messages, list):
for message in input_messages:
if isinstance(message, dict) and isinstance(message.get("content"), str):
return message["content"]
return ""
def _kwargs(self, value: Any) -> list[dict[str, Any]]:
if not isinstance(value, list):
return []
parsed_kwargs = []
for item in value:
if isinstance(item, dict):
parsed_kwargs.append(item)
continue
if isinstance(item, str):
try:
parsed = json.loads(item)
except json.JSONDecodeError:
parsed = {}
parsed_kwargs.append(parsed if isinstance(parsed, dict) else {})
continue
parsed_kwargs.append({})
return parsed_kwargs
def _instruction_ids(self, value: Any) -> list[str]:
if not isinstance(value, list):
return []
return [str(item) for item in value]
def _sample_id(self, canonical_name: str, record: dict[str, Any], row_index: int) -> str:
value = record.get("id")
if value in (None, ""):
value = row_index
return f"{canonical_name.replace('.', '_')}-{value}"
def _string_or_empty(self, value: Any) -> str:
if value in (None, ""):
return ""
return str(value)
def instruction_ids_from_sample(self, sample: Sample) -> list[str]:
return self._instruction_ids(sample.metadata.get("instruction_id_list"))
def kwargs_from_sample(self, sample: Sample) -> list[dict[str, Any]]:
return self._kwargs(sample.metadata.get("kwargs"))
def prompt_key_from_sample(self, sample: Sample) -> str:
return self._string_or_empty(sample.metadata.get("key"))
class MATH500JsonlLoader(BaseLoader):
def iter_raw_records(self, item: ResolvedRegistryItem) -> Iterator[RawRecord]:
with item.path.open("r", encoding="utf-8") as handle:
for row_index, line in enumerate(handle, start=1):
if not line.strip():
continue
yield RawRecord(
source_path=str(item.path),
source_file=item.entry.canonical_name,
row_index=row_index,
record=json.loads(line),
)
def iter_samples(self, item: ResolvedRegistryItem) -> Iterator[Sample]:
for raw_record in self.iter_raw_records(item):
record = raw_record.record
metadata = self._metadata(record, raw_record.row_index)
yield Sample(
sample_id=self._sample_id(item.entry.canonical_name, record, raw_record.row_index),
source_file=item.entry.canonical_name,
source_path=str(item.path),
module=item.entry.module,
subset=item.entry.subset,
task=item.entry.task,
dimension=self._string_or_none(record.get("subset_key")) or "default",
prompt=self._prompt(record),
reference={"target": self._string_or_empty(record.get("target"))},
metadata=metadata,
raw_record=record,
)
def _metadata(self, record: dict[str, Any], row_index: int) -> dict[str, Any]:
metadata: dict[str, Any] = {}
raw_metadata = record.get("metadata")
if isinstance(raw_metadata, dict):
metadata.update(raw_metadata)
metadata["subset_key"] = self._string_or_none(record.get("subset_key"))
metadata["row_index"] = row_index
return metadata
def _prompt(self, record: dict[str, Any]) -> str:
input_messages = record.get("input")
if isinstance(input_messages, list):
for message in input_messages:
if isinstance(message, dict) and isinstance(message.get("content"), str):
return message["content"]
return ""
def _sample_id(self, canonical_name: str, record: dict[str, Any], row_index: int) -> str:
value = record.get("id")
subset_key = self._string_or_none(record.get("subset_key")) or "default"
if value in (None, ""):
value = row_index
subset_slug = re.sub(r"[^A-Za-z0-9]+", "_", subset_key).strip("_").lower()
return f"{canonical_name.replace('.', '_')}-{subset_slug}-{value}"
def _string_or_none(self, value: Any) -> str | None:
if value in (None, ""):
return None
return str(value)
def _string_or_empty(self, value: Any) -> str:
if value in (None, ""):
return ""
return str(value)
def expected_answer_from_sample(self, sample: Sample) -> str:
reference = sample.reference or {}
if isinstance(reference, dict):
value = reference.get("target")
else:
value = reference
return "" if value in (None, "") else str(value)
class AIMEJsonlLoader(MATH500JsonlLoader):
def iter_samples(self, item: ResolvedRegistryItem) -> Iterator[Sample]:
for raw_record in self.iter_raw_records(item):
record = raw_record.record
metadata = self._metadata(record, raw_record.row_index)
yield Sample(
sample_id=self._sample_id(item.entry.canonical_name, record, raw_record.row_index),
source_file=item.entry.canonical_name,
source_path=str(item.path),
module=item.entry.module,
subset=item.entry.subset,
task=item.entry.task,
dimension="default",
prompt=self._prompt(record),
reference={"target": self._string_or_empty(record.get("target"))},
metadata=metadata,
raw_record=record,
)
|