Add benchmark metadata and verified audio (upload in progress)

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+rubric_qa/rubric_qa.jsonl filter=lfs diff=lfs merge=lfs -text
+rubric_qa/rubric_qa_grouped.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,151 @@
 ---
-license: apache-2.0
+pretty_name: FineAudio-Judge Benchmark
+language:
+  - en
+  - zh
+task_categories:
+  - audio-classification
+tags:
+  - audio
+configs:
+  - config_name: rubric_qa
+    default: true
+    data_files:
+      - split: test
+        path: rubric_qa/rubric_qa.jsonl
+  - config_name: rubric_qa_grouped
+    data_files:
+      - split: test
+        path: rubric_qa/rubric_qa_grouped.jsonl
+  - config_name: preference_candidates
+    data_files:
+      - split: test
+        path: human_preference/candidate_groups.jsonl
+  - config_name: preference_rubrics
+    data_files:
+      - split: test
+        path: human_preference/prompt_rubrics.jsonl
+  - config_name: preference_pairwise
+    data_files:
+      - split: test
+        path: human_preference/human_preference_pairwise.jsonl
+  - config_name: preference_ranking
+    data_files:
+      - split: test
+        path: human_preference/human_preference_ranking.jsonl
+  - config_name: preference_annotations
+    data_files:
+      - split: test
+        path: human_preference/human_preference_annotations.jsonl
 ---
+
+# FineAudio-Judge Benchmark
+
+FineAudio-Judge Benchmark evaluates instruction following in complex text-to-audio generation through rubric-level judgments and agreement with human preferences. It contains two evaluation components:
+
+| Component | Main content |
+|---|---|
+| [Rubric-QA](rubric_qa/README.md) | 700 prompt-audio samples, 33,498 rubric questions, and reference yes/no labels obtained by majority vote over five model annotations |
+| [Human Preference](human_preference/README.md) | 100 prompts, each paired with three generated audio candidates (300 in total), together with prompt-derived rubrics and preference annotations from seven human annotators |
+
+All published audio is WAV. JSONL files use repository-relative audio paths.
+See the README in each component directory for field definitions, label
+provenance, subset names, consensus rules, and usage notes.
+
+## Dataset structure
+
+The benchmark is organized into two components, each with its own data files,
+audio directory, and README.
+
+```text
+.
+├── README.md
+├── rubric_qa/
+│   ├── README.md
+│   ├── rubric_qa.jsonl
+│   ├── rubric_qa_grouped.jsonl
+│   └── audio/
+└── human_preference/
+    ├── README.md
+    ├── candidate_groups.jsonl
+    ├── prompt_rubrics.jsonl
+    ├── human_preference_pairwise.jsonl
+    ├── human_preference_ranking.jsonl
+    ├── human_preference_annotations.jsonl
+    └── audio/
+```
+
+See the [Rubric-QA README](rubric_qa/README.md) and
+[Human Preference README](human_preference/README.md) for file descriptions
+and evaluation details.
+
+## Quick start
+
+Install the download client:
+
+```bash
+pip install huggingface_hub
+```
+
+Download the full benchmark:
+
+```python
+from huggingface_hub import snapshot_download
+
+dataset_root = snapshot_download(
+    repo_id="Zixchen/FineAudio-Judge-Benchmark",
+    repo_type="dataset",
+)
+
+print(dataset_root)
+```
+
+The downloaded repository contains both the JSONL records and the referenced
+WAV files. Audio paths in the JSONL records are relative to `dataset_root`.
+
+## License / 数据许可声明
+
+本数据集以下文件夹内的数据采用
+[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) 许可发布。
+
+The data in the following folders is released under the
+[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.
+
+- `rubric_qa/audio/a2/`
+- `rubric_qa/audio/a3/`
+- `rubric_qa/audio/cross_a1/`
+- `rubric_qa/audio/cross_a2/`
+- `rubric_qa/audio/qwen_audio_ood/`
+- `rubric_qa/audio/seedaudio_ood/`
+- `human_preference/audio/`
+
+本数据集 `rubric_qa/audio/a1/` 文件夹内的数据筛选自开源的
+FineVideo 数据集，使用时请参阅并遵守 FineVideo 的
+[Terms of Use](https://huggingface.co/datasets/HuggingFaceFV/finevideo#terms-of-use-for-finevideo)。
+
+The data in the `rubric_qa/audio/a1/` folder is derived from a filtered
+subset of the FineVideo dataset; please refer to and comply with the
+[Terms of Use of FineVideo](https://huggingface.co/datasets/HuggingFaceFV/finevideo#terms-of-use-for-finevideo)
+when using it.
+
+本项目新增的评测问题及标注采用 CC BY-NC 4.0 许可发布。
+该许可仅适用于本项目拥有或有权授予许可的部分，
+不替代第三方素材及相关模型或服务的适用许可和使用条款。
+
+Project-created evaluation questions and annotations are released under
+CC BY-NC 4.0. This license applies only to rights held or licensable by
+this project and does not replace the applicable licenses and terms of
+third-party materials or the relevant models and services.
+
+## Disclaimer / 免责声明
+
+部分数据来源于第三方开源数据集，不代表本项目的立场或观点。
+您对相关内容的复制、使用与分发须严格遵循开源数据集的许可条件，
+不得超出许可范围，并遵守适用的法律法规，自行承担相应责���。
+
+Some data is sourced from third-party open-source datasets and does not
+represent the views or opinions of this project. Your copying, use, and
+distribution of such content must strictly comply with the license terms
+of the source datasets and must not exceed the scope of those licenses.
+You are also responsible for complying with all applicable laws and
+regulations, and you assume the associated risks and liabilities.

human_preference/README.md

@@ -0,0 +1,105 @@
+# 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.
+

human_preference/audio/seedaudio_pref_006/variant_03.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d9a0441f7346997fa8c33a75f7d11da6bf7bc106dce6a0ced2c4f825969324a
+size 1994924

human_preference/audio/seedaudio_pref_007/variant_02.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ae2423fe2fc368fc0f12d3985186fed7ad6b6911a0b6d3866808cf82d4fcc50
+size 3146924

human_preference/audio/seedaudio_pref_007/variant_03.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ae66b3664778b3883ebb4cebaedf8728a628b768777b5f5e175a550afe5e1e1
+size 2784044

human_preference/audio/seedaudio_pref_008/variant_01.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc46f688775b76210abc9209f5ec0dce32cd2c366f363f863bfab5112a95e66f
+size 3127724

human_preference/audio/seedaudio_pref_008/variant_02.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8c5fb3f025aa036ee0cb0f623699fc085afbb2e5e48ad8ec6db531824c818c9
+size 2810924

rubric_qa/README.md

@@ -0,0 +1,133 @@
+# Rubric QA
+
+![QA subset topology](qa_subset_overview.png)
+
+This component contains 700 prompt-audio samples, split evenly across seven
+subsets, and 33,498 audible-only rubric questions.
+
+## Subsets
+
+| Subset | Audio samples | Rubric questions | Description |
+|---|---:|---:|---|
+| `a1` | 100 | 5,204 | Real-world audio derived from FineVideo videos |
+| `a2` | 100 | 2,123 | Deterministically composed audio scenes built from curated audio assets |
+| `a3` | 100 | 6,826 | Audio generated by Audio-Oscar from ASG-Bench scene prompts |
+| `cross_a1` | 100 | 5,204 | Audio generated by Audio-Oscar using prompts from A1 |
+| `cross_a2` | 100 | 2,123 | Audio generated by Audio-Oscar using prompts from A2 |
+| `qwen_audio_ood` | 100 | 6,009 | Audio generated by Qwen-Audio-3.0-Gen-Preview from ASG-Bench-derived scene prompts |
+| `seedaudio_ood` | 100 | 6,009 | Audio generated by SeedAudio 1.0 from ASG-Bench-derived scene prompts |
+| **Total** | **700** | **33,498** | |
+  
+OOD denotes out-of-distribution evaluation relative to the three training-data
+construction routes: real-world audio, deterministic composition, and
+prompt-driven Audio-Oscar generation. The two OOD subsets use the same set
+of ASG-Bench-derived scene prompts, with audio generated by
+Qwen-Audio-3.0-Gen-Preview and SeedAudio 1.0, respectively, and share the
+corresponding evaluation rubric sets.
+
+`rubric_set_id` identifies samples that reuse the same prompt-derived rubric
+set, such as `a1` and `cross_a1`, or the two generator-OOD subsets.
+
+## Consistency coverage
+
+The benchmark includes controlled prompt-side and audio-side counterfactual
+construction to cover varying degrees of prompt–audio consistency. Each of
+the seven subsets contains both `yes` and `no` reference labels.
+
+Across all 33,498 questions, 23,285 reference labels are `yes` (69.51%) and
+10,213 are `no` (30.49%). These are majority-vote reference-label statistics.
+
+| Subset | Yes labels | No labels | Yes (%) | No (%) |
+|---|---:|---:|---:|---:|
+| `a1` | 4,036 | 1,168 | 77.56 | 22.44 |
+| `a2` | 1,319 | 804 | 62.13 | 37.87 |
+| `a3` | 4,835 | 1,991 | 70.83 | 29.17 |
+| `cross_a1` | 3,304 | 1,900 | 63.49 | 36.51 |
+| `cross_a2` | 1,326 | 797 | 62.46 | 37.54 |
+| `qwen_audio_ood` | 4,041 | 1,968 | 67.25 | 32.75 |
+| `seedaudio_ood` | 4,424 | 1,585 | 73.62 | 26.38 |
+| **Total** | **23,285** | **10,213** | **69.51** | **30.49** |
+
+At the audio-sample level, the proportion of `yes` reference labels ranges
+from 5.77% to 100%.
+The benchmark therefore covers different degrees of agreement between
+prompt requirements and audio content.
+
+## `rubric_qa.jsonl` (question-level evaluation format)
+
+This file contains 33,498 audio–rubric question records used for the
+Rubric-Level Judgment evaluation. Each row contains one question about one
+audio sample and its reference yes/no label. Inference and accuracy
+calculation are performed at the question level.
+
+| Field | Description |
+|---|---|
+| `sample_id` | Audio-sample identifier; repeated across questions for the same audio sample |
+| `subset` | One of the seven subset names above |
+| `rubric_set_id` | Identifier linking samples that share the same rubric set |
+| `language` | Language code: `en` or `zh` |
+| `audio` | WAV path relative to the dataset repository root |
+| `rubric_id` | Rubric identifier within a sample; together with `sample_id`, identifies one question record |
+| `tuple_id` | Identifier of the atomic tuple associated with the rubric |
+| `subtype` | Rubric subtype, such as `exists`, `before`, or `overlap` |
+| `axis` | Evaluation axis: `content_fidelity`, `acoustic_presentation_fidelity`, or `temporal_structural_fidelity` |
+| `scope` | Rubric scope: `local`, `relational`, or `global` |
+| `domain` | List of domain tags, such as `speech`, `sound`, or `music` |
+| `parent_ids` | Parent rubric identifiers recording dependencies |
+| `question` | Binary question to be answered from the audio |
+| `votes` | Five model annotations, each recording `model`, `pass`, and `answer` |
+| `yes_votes` / `no_votes` | Numbers of annotations answering `yes` or `no` |
+| `vote_pattern` | Majority-to-minority vote counts: `5-0`, `4-1`, or `3-2` |
+| `label` | Majority-vote reference answer: `yes` or `no` |
+
+## `rubric_qa_grouped.jsonl` (grouped convenience view)
+
+This file contains 700 audio–prompt pairs. Each record includes an audio
+path, the full reference `prompt`, the rubric questions derived from
+that prompt, and their model annotations and majority-vote reference labels.
+
+`rubric_qa.jsonl` contains one audio–question pair per row, while this
+file presents each audio–prompt pair with all its rubric questions in
+a single record. Rubric-Level Judgment evaluation uses `rubric_qa.jsonl`
+for single-question inference.
+
+| Field | Description |
+|---|---|
+| `sample_id` | Unique audio-sample identifier |
+| `subset` | One of the seven subset names above |
+| `rubric_set_id` | Identifier linking samples that share the same rubric set |
+| `language` | Language code: `en` or `zh` |
+| `audio` | WAV path relative to the dataset repository root |
+| `prompt` | Full reference text from which the sample's rubric questions were derived |
+| `rubric_count` | Number of questions in the sample |
+| `yes_rubric_count` / `no_rubric_count` | Numbers of questions with majority-vote reference labels of `yes` or `no` |
+| `yes_rubric_ratio` / `no_rubric_ratio` | Corresponding reference-label counts divided by `rubric_count`, rounded to two decimal places |
+| `rubrics` | List of question objects containing `rubric_id`, `tuple_id`, `subtype`, `axis`, `scope`, `domain`, `parent_ids`, `question`, `votes`, `yes_votes`, `no_votes`, `vote_pattern`, and `label` |
+
+The yes/no ratios describe the reference-label distribution within each
+sample.
+
+## Reference labels
+
+Each audio–question pair has five model annotations: two runs of
+Gemini 2.5 Pro and three runs of Qwen3.5-Omni-Plus. The reference `label`
+is determined by majority vote and is either `yes` or `no`.
+
+The `votes` field preserves the individual annotations. `yes_votes`
+and `no_votes` record the vote counts for each answer. `vote_pattern`
+records the majority-to-minority vote counts: `5-0`, `4-1`, or `3-2`.
+
+These are model-derived reference labels. All five annotations and
+the resulting majority label are provided in both JSONL representations.
+
+## Evaluation protocol
+
+For Rubric-Level Judgment, each audio–question pair is evaluated
+independently with a yes/no prediction. Accuracy is computed across
+individual questions by comparing predictions with the majority-vote
+reference labels. Runtime batching may process multiple independent requests together for
+throughput.
+
+## `audio/`
+
+Audio is organized by the original public subset names.

rubric_qa/audio/a1/a1_finevideo_001.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:932cc14b5eade89c0f9663849ea7fa06f89fc07e3ab5c6cf5ebb29cdf0616593
+size 321070

rubric_qa/rubric_qa.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0563bba8d5d29aae3eec7c84363ca3578c7a589614c9a0844cad4ff5618b803
+size 24199217

rubric_qa/rubric_qa_grouped.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42bb326b70e57b63a75e46efbba6ef69a00a3a2c2c5ba7dcf7b614b5bd07c23c
+size 19672735