Update dataset card for AudioEval

README.md

@@ -1,7 +1,20 @@
 ---
 pretty_name: AudioEval
+license: other
 size_categories:
 - 1K<n<10K
+source_datasets:
+- original
+annotations_creators:
+- expert-generated
+- crowdsourced
+tags:
+- audio
+- text-to-audio
+- benchmark
+- evaluation
+- human-ratings
+- multimodal
 configs:
 - config_name: default
   data_files:
@@ -16,50 +29,165 @@ configs:
 
 # AudioEval
 
-AudioEval is a text-to-audio evaluation benchmark packaged for a private Hugging Face dataset repository.
-This release package keeps the audio files and split metadata in an `AudioFolder` layout so the dataset can be previewed on the Hub and later switched to public visibility without repacking.
+AudioEval is a benchmark for evaluating text-to-audio generation systems from two listener perspectives (`common` and `pro`) and five perceptual dimensions: `production_complexity`, `content_enjoyment`, `production_quality`, `textual_alignment`, and `content_usefulness`.
+This repository is packaged in an `AudioFolder`-compatible layout so it can be previewed directly on the Hugging Face Hub and later switched from private to public visibility without repacking.
 
-## What is included
+## Dataset Summary
 
-- `data/train`, `data/validation`, `data/test`: split directories with `.wav` files and `metadata.jsonl`.
-- `annotations/ratings.csv`: anonymized per-rater scores with stable `rater_id` values.
-- `annotations/system_info.csv`: mapping from `system_id` to the evaluated model name.
-- `annotations/prompts.tsv`: prompt metadata used to generate the audio clips.
-- `stats/*.csv`: reliability and model summary tables from the local analysis pipeline.
+- 4200 generated audio clips from 24 text-to-audio systems and 451 prompts.
+- 11.712 total hours of audio, with an average clip duration of 10.039 seconds.
+- Official splits: train=3360, validation=420, test=420.
+- 25200 per-rater annotation rows in `annotations/ratings.csv`.
+- 126000 dimension-level human scores overall, because each annotation row contains five dimension scores.
+- 12 raters in total: common=9, pro=3.
+- Rating rows by rater type: common=12600, pro=12600.
+- Prompts span five coarse scene categories: daily life, art and cultural, natural and outdoor, work and production, and transportation and travel.
 
-## Dataset size
+## What Is In This Repository
 
-- 3360 train clips, 420 validation clips, and 420 test clips.
-- 4200 total audio clips from 24 text-to-audio systems and 451 prompts.
-- 11.712 total hours of audio.
-- 25200 total ratings from 12 raters.
-- Rating rows by rater type: common=12600, pro=12600.
-- Unique raters by type: common=9, pro=3.
+- `data/train`, `data/validation`, `data/test`: split directories containing `.wav` files and a `metadata.jsonl` file for each split.
+- `annotations/ratings.csv`: anonymized per-rater annotations.
+- `annotations/prompts.tsv`: prompt text plus prompt-side metadata such as `scene_category` and `audioset_ontology`.
+- `annotations/system_info.csv`: mapping from `system_id` to the evaluated system name.
+- `annotations/release_summary.json`: summary counts for this packaged release.
+- `stats/*.csv`: analysis tables such as ICC, Krippendorff's alpha, significance tests, and model result summaries.
+
+## Suggested Uses
+
+- Benchmarking automatic evaluators for text-to-audio generation.
+- Training clip-level regression or distribution-prediction models from prompt-audio pairs.
+- Studying rating differences between professional and non-professional listeners.
+- Analyzing inter-rater reliability, disagreement, and cross-dimension correlations.
+
+## Split Sizes
+
+| Split | Clips |
+| --- | ---: |
+| train | 3360 |
+| validation | 420 |
+| test | 420 |
+
+## Data Structure
+
+### Split-Level Metadata
+
+Each row in `data/*/metadata.jsonl` corresponds to one generated audio clip.
+
+| Field | Description |
+| --- | --- |
+| `file_name` | Relative path to the audio file inside the split directory. |
+| `wav_name` | Original clip filename. |
+| `split` | One of `train`, `validation`, or `test`. |
+| `prompt_id` | Prompt identifier such as `P0004` or `N001`. |
+| `prompt_text` | Natural language prompt used to generate the clip. |
+| `scene_category` | Coarse manually assigned prompt category. |
+| `sound_event_count` | Number of sound events recorded in the prompt metadata. |
+| `audioset_ontology` | Prompt-side ontology category. |
+| `system_id` | System identifier such as `S001`. |
+| `system_name` | Human-readable system name. |
+| `num_ratings_common` | Number of common-rater judgments for the clip. |
+| `num_ratings_pro` | Number of professional-rater judgments for the clip. |
+| `num_ratings_overall` | Total number of judgments for the clip. |
+| `common_*_mean` | Mean score from common raters for a given dimension. |
+| `pro_*_mean` | Mean score from professional raters for a given dimension. |
+| `common_*_raw_scores` | List of raw common-rater scores for a given dimension. |
+| `pro_*_raw_scores` | List of raw professional-rater scores for a given dimension. |
+
+The five dimensions are `production_complexity`, `content_enjoyment`, `production_quality`, `textual_alignment`, and `content_usefulness`.
+
+### Per-Rater Annotations
+
+Each row in `annotations/ratings.csv` contains one rater's five-dimensional judgment for one clip.
+
+| Field | Description |
+| --- | --- |
+| `wav_name` | Clip filename. |
+| `split` | Dataset split for the clip. |
+| `prompt_id` | Prompt identifier. |
+| `system_id` | System identifier. |
+| `rater_type` | `common` or `pro`. |
+| `rater_id` | Stable anonymized rater identifier within this release. |
+| `production_complexity` | Integer score from 1 to 10. |
+| `content_enjoyment` | Integer score from 1 to 10. |
+| `production_quality` | Integer score from 1 to 10. |
+| `textual_alignment` | Integer score from 1 to 10. |
+| `content_usefulness` | Integer score from 1 to 10. |
+
+## Loading the Data
+
+Once you have access to the repository on the Hub, you can load the split metadata and audio files with `datasets`:
+
+```python
+from datasets import load_dataset
+
+train = load_dataset("Hui519/AudioEval", split="train")
+print(train[0]["audio"])
+print(train[0]["prompt_text"])
+```
+
+The anonymized annotation table can be loaded separately from `annotations/ratings.csv`.
+
+## Dataset Creation
+
+AudioEval is a benchmark release of machine-generated audio, not a collection of source recordings.
+Each clip is paired with its generation prompt, generating system identifier, and human evaluation results.
+This packaged Hub release was assembled from locally processed benchmark assets by:
+
+- standardizing prompt metadata into `annotations/prompts.tsv`,
+- merging raw MOS tables into a unified annotation table,
+- exporting split-level clip metadata into `metadata.jsonl`,
+- and anonymizing rater identities before publication.
+
+## Annotation Process
+
+- The release contains two listener groups: `common` and `pro`.
+- There are 9 common raters and 3 professional raters.
+- Each clip receives 3 common-rater and 3 professional-rater judgments for every dimension.
+- The five dimensions are production complexity, content enjoyment, production quality, textual alignment, and content usefulness.
+- Score values are integers from 1 to 10.
+
+## Biases, Risks, and Limitations
+
+- Prompt coverage is broad but not uniform across scene categories or sound-event complexity.
+- The prompts are written in English, but generated audio may contain non-speech sounds, speech-like content, music, or multilingual vocal content.
+- The benchmark is designed for evaluation research on text-to-audio systems and should not be treated as a general-purpose audio understanding dataset.
+- Because the audio is machine-generated, outputs may contain artifacts, distorted speech, or unsafe-sounding events that require additional filtering for downstream applications.
+- Anonymized `rater_id` values are release-specific identifiers and should not be linked back to source identities.
+
+## Privacy and Sensitive Information
+
+- This release intentionally excludes the original rater demographic tables.
+- No direct personally identifying information is included in the published files.
+- `annotations/ratings.csv` keeps only anonymized rater IDs, rater type, system ID, prompt ID, split, and scores.
+- Although the audio is machine-generated, some clips may depict or imitate human speech, children, alarms, animals, vehicles, or other safety-relevant scenarios.
+
+## License
 
-## Main split metadata fields
+The repository is currently private.
+The YAML metadata uses `license: other` because the final public release license has not been fixed yet.
+Until a public license is selected, access and reuse are governed by the repository owner's access settings and release terms.
+Before switching the repository to public visibility, replace this section and the YAML metadata with the final license identifier if you choose a standard Hugging Face-supported license.
 
-- `file_name`: relative path to the audio file inside each split directory.
-- `prompt_id`, `prompt_text`, `scene_category`, `sound_event_count`, `audioset_ontology`: prompt-side metadata.
-- `system_id`, `system_name`: evaluated generation system metadata.
-- `common_*_mean`, `pro_*_mean`: mean scores from common and professional raters.
-- `common_*_raw_scores`, `pro_*_raw_scores`: raw score lists per clip for each rating axis.
+## Citation
 
-The five rating axes are:
+If you use AudioEval, cite the paper below and, when relevant, the Hugging Face dataset repository.
 
-- `production_complexity`
-- `content_enjoyment`
-- `production_quality`
-- `textual_alignment`
-- `content_usefulness`
+Paper:
 
-## Privacy and release status
+- Hui Wang, Jinghua Zhao, Junyang Cheng, Cheng Liu, Yuhang Jia, Haoqin Sun, Jiaming Zhou, and Yong Qin. *AudioEval: Automatic Dual-Perspective and Multi-Dimensional Evaluation of Text-to-Audio-Generation*. arXiv:2510.14570, 2025. DOI: 10.48550/arXiv.2510.14570.
 
-- This package is intended for a private Hugging Face dataset repo first.
-- Rater demographic tables are intentionally omitted from this release package.
-- Per-rater annotations are anonymized before export.
-- Update the license and citation sections before switching the repo to public visibility.
+```bibtex
+@article{wang2025audioeval,
+  title={AudioEval: Automatic Dual-Perspective and Multi-Dimensional Evaluation of Text-to-Audio-Generation},
+  author={Wang, Hui and Zhao, Jinghua and Cheng, Junyang and Liu, Cheng and Jia, Yuhang and Sun, Haoqin and Zhou, Jiaming and Qin, Yong},
+  journal={arXiv preprint arXiv:2510.14570},
+  year={2025},
+  doi={10.48550/arXiv.2510.14570}
+}
+```
 
-## Notes
+## Release Status
 
 - The original source directory in this workspace remains unchanged.
 - The export directory can be regenerated by re-running `scripts/prepare_hf_audioeval_dataset.py`.
+- This card is written so the same package can remain useful both while private and after a later public release.