| 35 | |
| 36 | @dataclass(slots=True) |
| 37 | class NormalizedResult: |
| 38 | status: str |
| 39 | runtime: float | None |
| 40 | benchmark_name: str |
| 41 | model_alias: str |
| 42 | arch: str = "" |
| 43 | workflow_name: str = "" |
| 44 | prompt_bundle_name: str = "" |
| 45 | difficulty: str = "" |
| 46 | category: str = "" |
| 47 | benchmark_target: str = "" |
| 48 | ip_addr: str = "" |
| 49 | source_file: str = "" |
| 50 | source_dir: str = "" |
| 51 | target: str = "" |
| 52 | run_index: int | None = None |
| 53 | commands: list[Any] = field(default_factory=list) |
| 54 | history: list[Any] = field(default_factory=list) |
| 55 | prompts: dict[str, Any] = field(default_factory=dict) |
| 56 | benchmark: dict[str, Any] = field(default_factory=dict) |
| 57 | model: dict[str, Any] = field(default_factory=dict) |
| 58 | raw: dict[str, Any] = field(default_factory=dict) |
| 59 | |
| 60 | def to_dict(self) -> dict[str, Any]: |
| 61 | return { |
| 62 | "status": self.status, |
| 63 | "runtime": self.runtime, |
| 64 | "benchmark_name": self.benchmark_name, |
| 65 | "model_alias": self.model_alias, |
| 66 | "arch": self.arch, |
| 67 | "workflow_name": self.workflow_name, |
| 68 | "prompt_bundle_name": self.prompt_bundle_name, |
| 69 | "difficulty": self.difficulty, |
| 70 | "category": self.category, |
| 71 | "benchmark_target": self.benchmark_target, |
| 72 | "ip_addr": self.ip_addr, |
| 73 | "source_file": self.source_file, |
| 74 | "source_dir": self.source_dir, |
| 75 | "target": self.target, |
| 76 | "run_index": self.run_index, |
| 77 | "commands": self.commands, |
| 78 | "history": self.history, |
| 79 | "prompts": self.prompts, |
| 80 | "benchmark": self.benchmark, |
| 81 | "model": self.model, |
| 82 | "raw": self.raw, |
| 83 | } |
| 84 | |
| 85 | |
| 86 | def load_normalized_results( |
no outgoing calls
no test coverage detected