File size: 9,471 Bytes
f4cc109 | 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 | """Explicit schemas for every artifact exchanged in FinReason Cup 2026 Task 3.
Implemented with the standard library only so that participants can validate a
submission without installing a validation framework. Every ``from_dict``
raises :class:`SchemaError` with a human-readable, ID-anchored message.
"""
from __future__ import annotations
import json
from dataclasses import asdict, dataclass, field
from typing import Any, Dict, List, Mapping, Optional
__all__ = [
"ANSWER_KEYS",
"DevelopmentExample",
"EvaluationResult",
"GoldAnswer",
"ParticipantPrediction",
"SchemaError",
"SubmissionMetadata",
"TestInput",
]
#: The two keys that make up a FinMR answer object, in canonical order.
ANSWER_KEYS = ("extracted_value", "calculated_value")
class SchemaError(ValueError):
"""Raised when a record does not conform to the Task 3 schema."""
def _require_mapping(obj: Any, what: str) -> Mapping[str, Any]:
if not isinstance(obj, Mapping):
raise SchemaError(f"{what} must be a JSON object, got {type(obj).__name__}")
return obj
def _require_str(obj: Mapping[str, Any], key: str, what: str, *, allow_number: bool = False) -> str:
if key not in obj or obj[key] is None:
raise SchemaError(f"{what} is missing required field '{key}'")
value = obj[key]
if isinstance(value, str):
return value
if allow_number and isinstance(value, (int, float)) and not isinstance(value, bool):
# Numbers are accepted but normalised to text: FinMR answers are strings
# such as "-1,284" whose formatting must survive a round-trip.
return str(value)
raise SchemaError(f"{what} field '{key}' must be a string, got {type(value).__name__}")
@dataclass(frozen=True)
class GoldAnswer:
"""The reference answer of one case."""
extracted_value: str
calculated_value: str
@classmethod
def from_dict(cls, obj: Any, what: str = "answer") -> "GoldAnswer":
obj = _require_mapping(obj, what)
return cls(
extracted_value=_require_str(obj, "extracted_value", what, allow_number=True),
calculated_value=_require_str(obj, "calculated_value", what, allow_number=True),
)
def to_dict(self) -> Dict[str, str]:
return {
"extracted_value": self.extracted_value,
"calculated_value": self.calculated_value,
}
def to_judge_text(self) -> str:
"""Render the answer the way the reference judge prompt expects to see it."""
return json.dumps(self.to_dict(), ensure_ascii=False)
@dataclass(frozen=True)
class DevelopmentExample:
"""One fully specified public development case (input + gold answer)."""
id: str
dqc_id: str
query: str
answer: GoldAnswer
source: Dict[str, Any] = field(default_factory=dict)
@classmethod
def from_dict(cls, obj: Any) -> "DevelopmentExample":
obj = _require_mapping(obj, "development example")
case_id = _require_str(obj, "id", "development example", allow_number=True)
what = f"development example {case_id}"
source = obj.get("source", {})
if not isinstance(source, Mapping):
raise SchemaError(f"{what} field 'source' must be a JSON object")
return cls(
id=case_id,
dqc_id=_require_str(obj, "dqc_id", what),
query=_require_str(obj, "query", what),
answer=GoldAnswer.from_dict(obj.get("answer"), f"{what} field 'answer'"),
source=dict(source),
)
def to_dict(self) -> Dict[str, Any]:
out: Dict[str, Any] = {
"id": self.id,
"dqc_id": self.dqc_id,
"query": self.query,
"answer": self.answer.to_dict(),
}
if self.source:
out["source"] = self.source
return out
def to_input(self) -> "TestInput":
return TestInput(id=self.id, dqc_id=self.dqc_id, query=self.query)
@dataclass(frozen=True)
class TestInput:
"""One case as it is handed to a participant system: no answer."""
id: str
dqc_id: str
query: str
#: Tells pytest this dataclass is not a test class despite its name.
__test__ = False
@classmethod
def from_dict(cls, obj: Any) -> "TestInput":
obj = _require_mapping(obj, "input record")
case_id = _require_str(obj, "id", "input record", allow_number=True)
what = f"input record {case_id}"
for forbidden in ("answer", "extracted_value", "calculated_value"):
if forbidden in obj:
raise SchemaError(f"{what} must not contain gold field '{forbidden}'")
return cls(
id=case_id,
dqc_id=_require_str(obj, "dqc_id", what),
query=_require_str(obj, "query", what),
)
def to_dict(self) -> Dict[str, str]:
return {"id": self.id, "dqc_id": self.dqc_id, "query": self.query}
@dataclass(frozen=True)
class ParticipantPrediction:
"""One line of ``predictions.jsonl``.
``raw_output`` is optional and is never used for scoring in the native
format; it exists so teams can keep an audit trail of what their system
actually emitted.
"""
id: str
extracted_value: str
calculated_value: str
raw_output: Optional[str] = None
#: Which file format this came from. "native" submissions are always judged on
#: the two parsed fields, so attaching `raw_output` for your own audit trail
#: can never change your score. Only the "lm-eval" adapter sets this to
#: "lm-eval", where the model's raw text IS the answer and must be judged as
#: such for structural errors to remain reachable.
source_format: str = "native"
@classmethod
def from_dict(cls, obj: Any) -> "ParticipantPrediction":
obj = _require_mapping(obj, "prediction")
case_id = _require_str(obj, "id", "prediction", allow_number=True)
what = f"prediction {case_id}"
raw = obj.get("raw_output")
if raw is not None and not isinstance(raw, str):
raise SchemaError(f"{what} field 'raw_output' must be a string when present")
return cls(
id=case_id,
extracted_value=_require_str(obj, "extracted_value", what, allow_number=True),
calculated_value=_require_str(obj, "calculated_value", what, allow_number=True),
raw_output=raw,
)
def to_dict(self, *, include_raw: bool = False) -> Dict[str, Any]:
out: Dict[str, Any] = {
"id": self.id,
"extracted_value": self.extracted_value,
"calculated_value": self.calculated_value,
}
if include_raw and self.raw_output is not None:
out["raw_output"] = self.raw_output
return out
def to_judge_text(self) -> str:
"""Canonical JSON text shown to the judge as ``pred_answer``."""
return json.dumps(
{
"extracted_value": self.extracted_value,
"calculated_value": self.calculated_value,
},
ensure_ascii=False,
)
def judge_text(self) -> str:
"""Exactly what the judge sees for this prediction.
For a native submission this is always the canonical two-key object built
from the declared fields - never `raw_output`, so an optional audit field
cannot change a score. For an lm-eval record the raw model text is the
answer, so it is judged verbatim.
"""
if self.source_format == "lm-eval" and self.raw_output is not None:
return self.raw_output
return self.to_judge_text()
@dataclass(frozen=True)
class EvaluationResult:
"""The case-level audit record written to ``case_results.jsonl``."""
id: str
gold: Dict[str, str]
prediction: Optional[Dict[str, str]]
label: Optional[str]
judge_model: str
prompt_version: str
raw_judge_output: Optional[str] = None
from_cache: bool = False
error: Optional[str] = None
attempts: int = 0
def to_dict(self) -> Dict[str, Any]:
return asdict(self)
@dataclass(frozen=True)
class SubmissionMetadata:
"""Optional ``metadata.json`` submitted alongside ``predictions.jsonl``."""
team_name: str
contact_email: str
method_name: str
system_description: str = ""
models_used: List[str] = field(default_factory=list)
uses_external_api: bool = False
submission_version: str = "final"
@classmethod
def from_dict(cls, obj: Any) -> "SubmissionMetadata":
obj = _require_mapping(obj, "submission metadata")
models = obj.get("models_used", [])
if isinstance(models, str):
models = [models]
if not isinstance(models, (list, tuple)):
raise SchemaError("submission metadata field 'models_used' must be a list")
return cls(
team_name=_require_str(obj, "team_name", "submission metadata"),
contact_email=_require_str(obj, "contact_email", "submission metadata"),
method_name=_require_str(obj, "method_name", "submission metadata"),
system_description=str(obj.get("system_description", "")),
models_used=[str(m) for m in models],
uses_external_api=bool(obj.get("uses_external_api", False)),
submission_version=str(obj.get("submission_version", "final")),
)
def to_dict(self) -> Dict[str, Any]:
return asdict(self)
|