MINIkio/MIV-Audio-Judgment

MIV Audio Judgment — Metadata

MIV provides audio judgment tasks and structured supervision inspired by JEV decision formats. It is an independent project with no official JEV affiliation or endorsement.

I assembled this dataset to develop my own audio judgment model, but hardware limitations prevented completion of the intended model development.

Audio is not included.

Data

The release contains 1,120,796 question/target tasks across five sources. The splits are train 508,005, validation 523,377, and test 89,414.

Source	Tasks	Source conditions
FMA	606,924	Metadata: CC BY 4.0
MTG-Jamendo	483,492	Metadata: CC BY-NC-SA 4.0
PVQD	3,234	Version 1: CC BY 4.0
SCHAEFFER	24,810	Dataset: CC BY 4.0
SongDescriber	2,336	Annotations: CC BY-SA 4.0

Generation

Questions and some caption-derived answers were generated or paraphrased using zai-org/GLM-5.3-Flash and deepseek-ai/DeepSeek-V4-Flash-0731 through DeepInfra. Source labels and rating distributions are retained where used, and provenance distinguishes them from teacher-derived text. Some positive-tag questions use a local template.

Schema and use

Fields	Meaning
id, source, source_id	Task and source identifiers
audio_ref	Relative source audio ID and crop bounds
question, options	Question and ordered response options
target_value, target_distribution, target_source	Supervision and its recorded origin
provenance	Source annotations and generation provenance
group_id, supervision_group_id, split	Related-task grouping and data split

question.criteria_json and provenance.source_labels_json are JSON-encoded structures. See schema.json, statistics.json, and SOURCE_ID_SCHEMES.json for details.

import json
from pathlib import Path

with Path("data/cc-by-4.0/fma/train.jsonl").open() as stream:
    row = json.loads(next(stream))

criteria = json.loads(row["question"]["criteria_json"]) if row["question"]["criteria_json"] else None
if row["target_distribution"] is not None:
    supervision = list(zip(row["options"], row["target_distribution"]))

Choice probabilities follow the explicit option order. Noul uses false/true; Score uses ordered scale levels; VALUE supplies text or a number. Keep target and source-annotation fields as supervision rather than model input context.

Several tasks can refer to the same audio, and source labels or generated text may contain errors. Respect the provided splits and prior-use information when evaluating; this release is not an untouched benchmark.

Licensing

This is a mixed-license release. MIV's independently authored or generated original captions are offered under Apache-2.0, only to the extent that MIV holds the rights in those independent contributions. Original source captions, tags, annotations, and content derived from them continue to follow their respective source licenses and conditions. Generation, rewriting, or reformatting does not remove source obligations.

The current JSONL schema has no field certifying a row as an independently licensed MIV caption. A teacher name or target_source value is not an Apache eligibility flag: do not treat every generated target, row, source partition, or the whole dataset as Apache-2.0. See LICENSE_SCOPE.md for the boundary and LICENSE-APACHE-2.0.txt for the Apache license text.

The source-specific NOTICE.md and SOURCE_CONDITIONS.md files remain applicable. Attribution, non-commercial and share-alike requirements are retained where required by the corresponding source; no single source's conditions are imposed on every other partition.