File size: 2,512 Bytes
f5864f8 | 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 | """Training schema; labels/provenance never participate in model rendering."""
import math
from typing import Any, Literal
from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, model_validator
from decision_schema import Question, options
QUESTION = TypeAdapter(Question)
KINDS = ("choice", "score", "noul")
MODELS = (
"github-copilot/gpt-6-luna",
"github-copilot/grok-4.7",
"github-copilot/gpt-6-sol",
)
def validate_question(value):
q = QUESTION.validate_python(value).model_dump(exclude_none=True)
if not q.get("instructions"):
raise ValueError("instructions required")
if q["type"] in {"choice", "score"} and len(q["criteria"]) < 2:
raise ValueError("training requires at least two candidates")
return q
def label_key(label, kind):
if kind == "noul":
if isinstance(label, bool):
return "true" if label else "false"
if label in ("true", "false"):
return label
raise ValueError("noul label must be a boolean or true/false string")
if isinstance(label, bool):
raise ValueError("boolean is not a choice/score label")
return str(label)
class Target(BaseModel):
model_config = ConfigDict(extra="forbid")
hard_label: str | int | bool
evidence: list[dict[str, Any]] = Field(min_length=1)
ambiguity: Literal["none", "ambiguous", "insufficient", "conflict"] = "none"
class Case(BaseModel):
model_config = ConfigDict(extra="forbid")
case_id: str
group_id: str
domain: str
task_family: str
language: Literal["zh", "en", "mixed"]
state: str | dict[str, Any] | list[Any]
questions: dict[str, dict[str, Any]] = Field(min_length=1, max_length=8)
targets: dict[str, Target]
@model_validator(mode="after")
def check_targets(self):
if set(self.questions) != set(self.targets):
raise ValueError("question/target IDs differ")
for qid, question in self.questions.items():
q = validate_question(question)
self.questions[qid] = q
key = label_key(self.targets[qid].hard_label, q["type"])
if key not in [k for k, _ in options(q)]:
raise ValueError("target is not a candidate")
return self
def check_distribution(values):
if not values or any(not math.isfinite(x) or x < 0 for x in values):
raise ValueError("invalid distribution")
if abs(sum(values) - 1) > 1e-5:
raise ValueError("probabilities must sum to one")
|