moondream2-coyo-5M-captions

## Moondream2 COYO-700M 5M subset captions

A 5-million image, text pair subset of COYO-700M dataset, captioned with Moondream2 (rev=`2024-05-08`). Captioning question is `Write a long caption for this image given the alt text: {alt_text}`.

### Sampling conditions

Randomly sampled from 5 million images from COYO-700M images that fit to the following filters:

```
filters = [
("width", ">=", 256),
("height", ">=", 256),
("aesthetic_score_laion_v2", ">=", 5.2),
("watermark_score", "<=", 0.40),
("clip_similarity_vitl14", ">=", 0.1),
]
```

## Moondream2 COYO-700M 5M 子集字幕数据集
该数据集为COYO-700M数据集的500万图像-文本对子集，由Moondream2（版本修订号=`2024-05-08`）生成字幕。用于生成字幕的提示指令为`Write a long caption for this image given the alt text: {alt_text}`，即`针对该图像，结合备选文本（alt text）撰写一段详细字幕：{alt_text}`。

### 采样条件
该子集从COYO-700M的500万张符合以下筛选条件的图像中随机采样获得：

筛选条件 = [
("图像宽度", ">=", 256),
("图像高度", ">=", 256),
("LAION V2美学评分", ">=", 5.2),
("水印评分", "<=", 0.40),
("CLIP（Contrastive Language-Image Pre-training）ViT-L/14相似度", ">=", 0.1),
]

搜集汇总

背景与挑战

背景概述

该数据集是从COYO-700M中随机抽取500万张图像构成的子集，每张图像均通过Moondream2模型生成长描述文本。采样时应用了严格的筛选条件，包括图像尺寸、美学评分、水印评分和CLIP相似度阈值。