AIME-survey

```
from datasets import load_dataset
dataset = load_dataset('disco-eth/AIME-survey')
```

# AIME Survey: AI Music Evaluation Dataset

This survey dataset accompanies the [AIME audio dataset](https://huggingface.co/datasets/disco-eth/AIME).

The AIME Survey dataset consists of 15,600 pairwise audio comparisons rated by more than 2,500 human participants regarding the music quality and text-audio alignment of 12 state-of-the-art music generation models (as of July 2024). The comparisons were made between 10 second snippets of the audio tracks.

The dataset contains the following fields:

- **question-type**: The type of question for the evaluation of the two audio tracks. This can be either 'Text-Audio Alignment' or 'Music Quality'
- **description**: The tag-based music description that was used to generate the tracks.
- **model-1**: The music generation model that generated track-1.
- **track-1-id**: The id for track-1. This corresponds to the id's in the AIME audio dataset.
- **track-1-begin**: The timestamp for the begin of the audio snippet from track-1.
- **track-1-end**: The timestamp for the end of the audio snippet from track-1.
- **model-2**: The music generation model that generated track-2.
- **track-2-id**: The id for track-2. This corresponds to the id's in the AIME audio dataset.
- **track-2-begin**: The timestamp for the begin of the audio snippet from track-2.
- **track-2-end**: The timestamp for the end of the audio snippet from track-2.
- **answer**: Whether the participant preferred the audio snippet from track-1 (answer=1) or track-2 (answer=2).

For more information or to cite our work please see [Benchmarking Music Generation Models and Metrics via Human Preference Studies](https://ieeexplore.ieee.org/abstract/document/10887745).

python
from datasets import load_dataset
dataset = load_dataset('disco-eth/AIME-survey')

# AIME 调查数据集：AI音乐评估数据集
本调查数据集配套于[AIME音频数据集](https://huggingface.co/datasets/disco-eth/AIME)。
AIME调查数据集包含15600组成对音频对比样本，由超过2500名人类参与者针对截至2024年7月的12款顶尖音乐生成模型的音乐质量与文本-音频对齐度进行评分。对比样本取自音频轨道的10秒节选片段。

本数据集包含以下字段：
- **问题类型（question-type）**：用于评估两段音频轨道的问题类别，可选值为「文本-音频对齐（Text-Audio Alignment）」或「音乐质量（Music Quality）」
- **描述信息（description）**：用于生成该音频轨道的基于标签的音乐描述文本
- **模型1（model-1）**：生成轨道1的音乐生成模型
- **轨道1编号（track-1-id）**：轨道1的唯一标识，与AIME音频数据集中的编号一一对应
- **轨道1起始时间戳（track-1-begin）**：轨道1所取音频片段的起始时间点
- **轨道1结束时间戳（track-1-end）**：轨道1所取音频片段的结束时间点
- **模型2（model-2）**：生成轨道2的音乐生成模型
- **轨道2编号（track-2-id）**：轨道2的唯一标识，与AIME音频数据集中的编号一一对应
- **轨道2起始时间戳（track-2-begin）**：轨道2所取音频片段的起始时间点
- **轨道2结束时间戳（track-2-end）**：轨道2所取音频片段的结束时间点
- **作答结果（answer）**：参与者更偏好轨道1的音频片段（answer=1）还是轨道2的音频片段（answer=2）

如需获取更多信息或引用本研究，请参阅《通过人类偏好研究评估音乐生成模型与指标》（Benchmarking Music Generation Models and Metrics via Human Preference Studies），链接：https://ieeexplore.ieee.org/abstract/document/10887745。