Orienter / evaluation /cache_manifest.json
stereoid's picture
Add files using upload-large-folder tool
6d35aff verified
Raw
History Blame Contribute Delete
1.59 kB
{
"schema_version": 1,
"artifact": "embedding_dict.json",
"distribution": "external_runtime_asset_not_tracked_by_git",
"embedding_model": "embedding-3",
"semantic_similarity_threshold": 0.85,
"semantic_similarity_operation": "raw_dot_product_historical_implementation_on_approximately_unit_norm_vectors",
"sha256": "df3c3dfd625f8205c278d9d7b22641cfbf351c886559abb1894c663bb657d067",
"size_bytes": 548127946,
"entry_count": 19644,
"embedding_dimension": 2048,
"non_finite_vector_count": 0,
"inconsistent_dimension_count": 0,
"blank_key_count": 1,
"vector_l2_norm": {
"minimum": 0.9999337429879811,
"mean": 0.999999939836742,
"median": 0.9999999995442519,
"maximum": 1.000054469389027,
"maximum_absolute_deviation_from_one": 0.0000662570120189
},
"frozen_at": "2026-08-22",
"recommended_file_mode": "0444",
"runtime_environment": {
"ORIENTER_EMBEDDING_CACHE": "/absolute/path/to/embedding_dict.json",
"ORIENTER_EMBEDDING_OFFLINE": "1",
"ORIENTER_EMBEDDING_READONLY": "1"
},
"notes": [
"Verify the SHA-256 before evaluation.",
"Offline mode fails before an embedding API call when a required category is absent.",
"The paper describes cosine similarity, while the historical evaluator uses a raw dot product. The cached vectors are approximately unit norm, so the operations are numerically close but not formally identical.",
"The frozen artifact is unchanged from the historical evaluator cache; the single blank key is retained for byte identity and is not used by the validated evaluation paths."
]
}