Datasets:
Formats:
json
Size:
1K - 10K
Tags:
audio
text
audio-language-model
audio-text-alignment
instruction-following
rubric-based-evaluation
Clarify benchmark overview, evaluation protocols, and dataset tags
Browse files- README.md +47 -1
- rubric_qa/README.md +2 -3
README.md
CHANGED
|
@@ -5,8 +5,16 @@ language:
|
|
| 5 |
- zh
|
| 6 |
task_categories:
|
| 7 |
- audio-classification
|
|
|
|
| 8 |
tags:
|
| 9 |
- audio
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
configs:
|
| 11 |
- config_name: rubric_qa
|
| 12 |
default: true
|
|
@@ -41,7 +49,20 @@ configs:
|
|
| 41 |
|
| 42 |
# FineAudio-Judge Benchmark
|
| 43 |
|
| 44 |
-
FineAudio-Judge
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
| Component | Main content |
|
| 47 |
|---|---|
|
|
@@ -52,6 +73,31 @@ All published audio is WAV. JSONL files use repository-relative audio paths.
|
|
| 52 |
See the README in each component directory for field definitions, label
|
| 53 |
provenance, subset names, consensus rules, and usage notes.
|
| 54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
## Dataset structure
|
| 56 |
|
| 57 |
The benchmark is organized into two components, each with its own data files,
|
|
|
|
| 5 |
- zh
|
| 6 |
task_categories:
|
| 7 |
- audio-classification
|
| 8 |
+
- audio-text-to-text
|
| 9 |
tags:
|
| 10 |
- audio
|
| 11 |
+
- text
|
| 12 |
+
- audio-language-model
|
| 13 |
+
- audio-text-alignment
|
| 14 |
+
- instruction-following
|
| 15 |
+
- rubric-based-evaluation
|
| 16 |
+
- human-preference
|
| 17 |
+
- benchmark
|
| 18 |
configs:
|
| 19 |
- config_name: rubric_qa
|
| 20 |
default: true
|
|
|
|
| 49 |
|
| 50 |
# FineAudio-Judge Benchmark
|
| 51 |
|
| 52 |
+
[Code](https://github.com/QwenAudio/FunResearch/tree/main/FineAudio-Judge) |
|
| 53 |
+
[Model](https://huggingface.co/Zixchen/FineAudio-Judge)
|
| 54 |
+
|
| 55 |
+
FineAudio-Judge is a fine-grained evaluation framework for assessing
|
| 56 |
+
whether audio matches text descriptions or generation instructions
|
| 57 |
+
for complex audio scenes. It structurally decomposes the text into
|
| 58 |
+
events, attributes, and inter-event relations, then converts them into
|
| 59 |
+
atomic binary rubric questions for audio-based judgments.
|
| 60 |
+
|
| 61 |
+
This repository provides the FineAudio-Judge evaluation benchmark.
|
| 62 |
+
It covers real, composed, and generated audio and supports evaluation
|
| 63 |
+
of rubric-level judgments and agreement with human preferences.
|
| 64 |
+
|
| 65 |
+
The benchmark contains two evaluation components:
|
| 66 |
|
| 67 |
| Component | Main content |
|
| 68 |
|---|---|
|
|
|
|
| 73 |
See the README in each component directory for field definitions, label
|
| 74 |
provenance, subset names, consensus rules, and usage notes.
|
| 75 |
|
| 76 |
+
The benchmark is organized by evaluation component and subset.
|
| 77 |
+
English and Chinese records are stored together, with the language
|
| 78 |
+
of each record indicated by the `language` field.
|
| 79 |
+
|
| 80 |
+
## Evaluation overview
|
| 81 |
+
|
| 82 |
+
### Rubric-QA
|
| 83 |
+
|
| 84 |
+
This component evaluates yes/no judgments at the individual rubric
|
| 85 |
+
level. Each audio–rubric question pair is evaluated independently.
|
| 86 |
+
The reference label is determined by majority vote across five model
|
| 87 |
+
annotations. Predictions are compared with these reference labels
|
| 88 |
+
to compute rubric-level accuracy and yes/no recall.
|
| 89 |
+
|
| 90 |
+
### Human Preference
|
| 91 |
+
|
| 92 |
+
Each candidate audio is scored by the proportion of rubric questions
|
| 93 |
+
answered yes. Scores are used to compare and rank candidates for the
|
| 94 |
+
same prompt, with accuracy measured against human pairwise
|
| 95 |
+
preferences and three-candidate rankings.
|
| 96 |
+
|
| 97 |
+
See the [Rubric-QA README](rubric_qa/README.md) and
|
| 98 |
+
[Human Preference README](human_preference/README.md)
|
| 99 |
+
for detailed evaluation protocols.
|
| 100 |
+
|
| 101 |
## Dataset structure
|
| 102 |
|
| 103 |
The benchmark is organized into two components, each with its own data files,
|
rubric_qa/README.md
CHANGED
|
@@ -30,9 +30,8 @@ set, such as `a1` and `cross_a1`, or the two generator-OOD subsets.
|
|
| 30 |
|
| 31 |
## Consistency coverage
|
| 32 |
|
| 33 |
-
The benchmark
|
| 34 |
-
|
| 35 |
-
the seven subsets contains both `yes` and `no` reference labels.
|
| 36 |
|
| 37 |
Across all 33,498 questions, 23,285 reference labels are `yes` (69.51%) and
|
| 38 |
10,213 are `no` (30.49%). These are majority-vote reference-label statistics.
|
|
|
|
| 30 |
|
| 31 |
## Consistency coverage
|
| 32 |
|
| 33 |
+
The benchmark covers varying degrees of audio–text consistency.
|
| 34 |
+
Each of the seven subsets contains both `yes` and `no` reference labels.
|
|
|
|
| 35 |
|
| 36 |
Across all 33,498 questions, 23,285 reference labels are `yes` (69.51%) and
|
| 37 |
10,213 are `no` (30.49%). These are majority-vote reference-label statistics.
|