Datasets:
Formats:
json
Size:
1K - 10K
Tags:
audio
text
audio-language-model
audio-text-alignment
instruction-following
rubric-based-evaluation
File size: 4,338 Bytes
dc760ae | 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 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 | # Human Preference
This component contains 100 prompts derived from ASG-Bench, each used to
independently generate three audio candidates with SeedAudio 1.0
(300 candidates in total), together with prompt-derived rubrics and
annotations from seven human annotators. All released audio files are in
WAV format.
The release contains all 100 prompt groups and their individual annotation
records. The paper evaluation uses 279 pairwise comparisons with a unique
winning candidate and 84 three-candidate ranking groups with a unique
modal ranking among valid complete annotations.
## Files
### `candidate_groups.jsonl`
One unlabeled group per prompt. Each row contains `group_id`, `language`,
`category`, `prompt`, and three nested candidates with `candidate_id`, `audio`,
and `duration_seconds`. It intentionally contains no preference label so the
groups can be independently re-annotated.
### `prompt_rubrics.jsonl`
One prompt and its nested rubric list per row. Rubric fields are `rubric_id`,
`question`, `axis`, `scope`, `subtype`, `domain`, `basis`, and `parent_ids`.
The rubric questions are produced from each prompt using the structured
rubric decomposition method proposed in FineAudio-Judge.
### `human_preference_annotations.jsonl`
This file contains 700 individual annotation records, seven per prompt
group. Annotators are identified by stable anonymous IDs
`annotator_01` through `annotator_07`.
Each row contains the annotation status, `unable_to_judge`, the
best-to-worst candidate `ranking`, and the candidate `presentation_order`.
### `human_preference_pairwise.jsonl`
This file contains 279 pairwise comparisons with a unique modal preference
among the 300 possible candidate pairs across 100 prompt groups, based on
the annotators' complete rankings.
`label` identifies the winning candidate. `votes` records the vote counts
for both candidates, together with unable-to-judge and missing or invalid
annotations. `vote_pattern` records `winner_votes:loser_votes`.
Consensus strata are:
| Level | Definition | Observed vote patterns | Comparisons |
|---|---|---|---:|
| `high` | loser votes = 0 | `7:0`, `6:0`, `5:0`, `4:0` | 68 |
| `medium` | loser votes = 1 | `6:1`, `5:1`, `4:1`, `3:1` | 91 |
| `low` | loser votes >= 2 | `5:2`, `4:3`, `4:2`, `3:2` | 120 |
Vote patterns are shown as `winner_votes:loser_votes`, with unable-to-judge
and missing or invalid annotations counted separately.
### `human_preference_ranking.jsonl`
This file contains the 84 three-candidate ranking groups used in the paper
evaluation. Each reference ranking is the unique mode among the valid
complete rankings submitted by the seven annotators. Groups with tied
modal rankings are excluded.
`label` gives the candidate order from best to worst. `vote_distribution`
records each observed complete ranking and its count. `consensus_level`
is `strict_majority` when the modal ranking receives at least four votes,
and `unique_plurality` when the unique mode receives two or three votes.
## Categories
| Category | Groups |
|---|---:|
| `Cinematic` | 18 |
| `Environmental` | 20 |
| `Field Reporting` | 19 |
| `Narrational` | 19 |
| `Podcast_Broadcast` | 24 |
## Evaluation protocol
For each candidate audio, the judge jointly evaluates the rubrics derived
from its prompt and predicts `yes` or `no` for each rubric. The proportion
of `yes` predictions forms the candidate's prompt–audio consistency score:
score = number of yes predictions / number of rubrics
Candidates generated from the same prompt are compared and ranked by
their consistency scores. Pairwise accuracy is evaluated against
`human_preference_pairwise.jsonl`. A three-candidate ranking is counted
as correct when the predicted order exactly matches the reference
order in `human_preference_ranking.jsonl`.
For pairwise evaluation, equal consistency scores for the two candidates
count as an incorrect judgment. For three-candidate ranking evaluation,
equal consistency scores for any two candidates count as an incorrect
ranking.
Evaluation is performed at the candidate-audio level, with multiple rubrics
for the same candidate evaluated jointly in each model request. Longer
rubric lists are divided into chunks of up to 80 questions by default.
Predictions from all chunks are combined into a single consistency score
for that candidate.
|