Matthew Ford
feat: expand substitution-triplet benchmark 8 -> 20
e6d699c
Raw
History Blame Contribute Delete
1.26 kB
{
"admission": {
"discordance_rule": "Morgan/Tanimoto(target,distractor) > Morgan/Tanimoto(target,preferred)",
"distractor_rule": "deterministic max-Tanimoto; same coarse odour class preferred, else global structural nearest; no perceptual label used in selection",
"identity_rule": "structural identity via registry CAS, or trade-name resolved to a parseable SMILES from the empirical corpus (trade names allowed per project decision)",
"preference_rule": "independent authoritative supplier substitution claim prefers substitute over distractor",
"required_fields": [
"target",
"preferred_substitute",
"structural_distractor",
"preference_provenance"
]
},
"admitted_triplet_count": 20,
"coverage_note": "Most supplier substitutions involve naturals or proprietary bases without a single-structure identity, so the admissible set is bounded by structural identity on both arms. 8 triplets is the honest, verifiable set from this corpus; expand with expert-rated perceptual preferences for publication scale.",
"label_access": "evaluation_only; forbidden in training and model selection",
"schema_version": 1,
"skipped": {
"no_distractor": 3,
"no_identity": 60
},
"status": "complete"
}