(
payload: dict[str, Any],
source_path: Path,
benchmark_index: dict[str, dict[str, Any]],
)
| 162 | |
| 163 | |
| 164 | def _normalize_task_result( |
| 165 | payload: dict[str, Any], |
| 166 | source_path: Path, |
| 167 | benchmark_index: dict[str, dict[str, Any]], |
| 168 | ) -> NormalizedResult: |
| 169 | details = payload.get("details", {}) |
| 170 | if not isinstance(details, dict): |
| 171 | details = {} |
| 172 | |
| 173 | benchmark_payload = _as_dict(details.get("benchmark")) |
| 174 | benchmark_name = str(payload.get("benchmark_name") or benchmark_payload.get("name") or "") |
| 175 | benchmark_metadata = _merge_benchmark_metadata(benchmark_name, benchmark_payload, benchmark_index) |
| 176 | |
| 177 | model_payload = _as_dict(details.get("model")) |
| 178 | model_alias = _canonicalize_model_alias( |
| 179 | payload.get("model_alias") or model_payload.get("alias") or model_payload.get("name") |
| 180 | ) |
| 181 | |
| 182 | return NormalizedResult( |
| 183 | status=str(payload.get("status", "")), |
| 184 | runtime=_float_or_none(payload.get("runtime")), |
| 185 | benchmark_name=benchmark_name or str(benchmark_metadata.get("name", "")), |
| 186 | model_alias=model_alias, |
| 187 | arch=_infer_arch(details, source_path), |
| 188 | workflow_name=_normalize_result_namespace(details.get("workflow_name", "")), |
| 189 | prompt_bundle_name=str(details.get("prompt_bundle_name", "")), |
| 190 | difficulty=str(benchmark_metadata.get("difficulty", "")), |
| 191 | category=str(benchmark_metadata.get("category", benchmark_metadata.get("type", ""))), |
| 192 | benchmark_target=str(benchmark_metadata.get("target", "")), |
| 193 | ip_addr=str(payload.get("ip_addr", "")), |
| 194 | source_file=str(source_path), |
| 195 | source_dir=str(source_path.parent), |
| 196 | target=str(details.get("target") or benchmark_metadata.get("target", "")), |
| 197 | run_index=_int_or_none(details.get("round_index")), |
| 198 | commands=_as_list(details.get("commands")), |
| 199 | history=_as_list(details.get("history")), |
| 200 | prompts=_as_dict(details.get("prompts")), |
| 201 | benchmark=benchmark_metadata, |
| 202 | model=model_payload, |
| 203 | raw=payload, |
| 204 | ) |
| 205 | |
| 206 | |
| 207 | def _normalize_legacy_result( |
no test coverage detected