Spaces:
Sleeping
Sleeping
File size: 3,423 Bytes
315cbe0 a969e99 315cbe0 a969e99 315cbe0 266f01b 315cbe0 a969e99 315cbe0 266f01b a969e99 | 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 | from dataclasses import dataclass
from typing import Literal, Optional
ResolutionStrategy = Literal["exact", "normalized", "fuzzy", "no_match"]
EntityType = Literal["model", "benchmark", "metric", "harness", "org"]
@dataclass
class ResolutionResult:
"""Outcome of one `Resolver.resolve` call.
Core matching fields (always populated):
- `raw_value`, `entity_type`, `source_config`: echo of the inputs
- `canonical_id`: the matched canonical (None on no_match). For
models with a `root_model_id` set, this is the IDENTITY ROOT —
i.e. the unquantized base — so callers reasoning about
same-identity quantizations get one canonical instead of N.
- `strategy`: which matcher fired (or "no_match")
- `confidence`: 0.0–1.0
Enrichment fields (populated when the `Resolver` is constructed
with a `CanonicalStore`; otherwise None):
- `review_status`: review state of the matched canonical
- `parent_canonical_id`: family/variant parent (for models, the
`variant` edge in `parents`; for benchmarks/orgs, the
`parent_*_id` scalar column).
- `resolved_leaf_id`: the originally-matched canonical before
any root-collapse. Equals `canonical_id` when no quantized
chain. Models only.
- `root_model_id`: identity root via quantized-only walk. NULL
when the matched leaf IS the root. Models only.
- `lineage_origin_org_id`: deepest non-variant ancestor's
org_id. Models only.
- `parents`: full typed-edge list of the matched leaf. Models only.
- `open_weights`: True/False/None. Models only.
- `release_date`: YYYY-MM or YYYY-MM-DD. Models only.
- `params_billions`: approximate parameter count. Models only.
- `family_key`: canonical_families.id this benchmark belongs to.
Defaults to the benchmark's own id for singleton families
(when no curated multi-benchmark family covers it). Benchmarks
only. See notes/hierarchy-alignment.md §3.
- `composite_keys`: canonical_composites.id values where this
benchmark appears (via the composite's source_configs ↔ EEE
folders chain). Benchmarks only; empty list when none.
- `category`: curated single-valued category from the family
(general / agentic / reasoning / knowledge / multimodal /
tool-use / math / security / factuality / reward-modelling /
safety / code / instruction-following / other). Benchmarks
only; None when no category curated.
"""
raw_value: str
entity_type: EntityType
source_config: Optional[str]
canonical_id: Optional[str]
strategy: ResolutionStrategy
confidence: float
# Enrichment fields — None when no CanonicalStore is attached.
review_status: Optional[str] = None
parent_canonical_id: Optional[str] = None
resolved_leaf_id: Optional[str] = None
root_model_id: Optional[str] = None
lineage_origin_org_id: Optional[str] = None
parents: Optional[list[dict]] = None
open_weights: Optional[bool] = None
release_date: Optional[str] = None
params_billions: Optional[float] = None
# Benchmark-only enrichment (see notes/hierarchy-alignment.md §3).
family_key: Optional[str] = None
composite_keys: Optional[list[str]] = None
category: Optional[str] = None
@dataclass
class ResolverConfig:
threshold: float = 0.85
|