sync-pilot / sync_pilot /taxonomy /model_eval_plan.py
Emre Sarigöl
Deploy sync_pilot dashboard - 2026-06-10 16:43
a01358a
Raw
History Blame Contribute Delete
2.01 kB
"""Sidecar model-evaluation design for culturally specific music tagging."""
from __future__ import annotations
from typing import Any
MODEL_CANDIDATES: list[dict[str, Any]] = [
{
"id": "laion/clap-htsat-unfused",
"role": "current_zero_shot_baseline",
"canonical_tags_mutated": False,
"notes": "Keep as candidate evidence until evaluated per taxonomy dimension.",
},
{
"id": "OpenMuQ/MuQ-MuLan-large",
"role": "zero_shot_music_text_candidate",
"canonical_tags_mutated": False,
"notes": "Most relevant drop-in zero-shot candidate; evaluate as sidecar first.",
},
{
"id": "m-a-p/MERT-v1-330M",
"role": "embedding_candidate",
"canonical_tags_mutated": False,
"notes": "Use embeddings for linear probes or nearest-neighbor review labels, not direct tags.",
},
{
"id": "OpenMuQ/MuQ-large-msd-iter",
"role": "embedding_candidate",
"canonical_tags_mutated": False,
"notes": "Alternative music foundation embedding model for supervised probes.",
},
]
EVALUATION_DIMENSIONS = [
"genre_family",
"genre_subtype",
"instrumentation",
"vocal_configuration",
"vocal_technique",
"mood",
"arrangement_aesthetic",
"recording_context",
]
METRICS = [
"top1_accuracy_for_single_select",
"top3_recall_for_review_candidates",
"precision_at_emitted_threshold",
"false_positive_examples",
"abstention_rate",
]
def model_eval_plan() -> dict[str, Any]:
return {
"canonical_tags_mutated": False,
"minimum_reviewed_tracks": 30,
"preferred_reviewed_tracks": 60,
"candidates": MODEL_CANDIDATES,
"dimensions": EVALUATION_DIMENSIONS,
"metrics": METRICS,
"decision_rule": (
"Promote a model per dimension only when it improves reviewed-label "
"precision/recall over CLAP and produces inspectable false-positive examples."
),
}