BLIP3o-Pretrain

## BLIP3o Pretrain Long-Caption Dataset

This collection contains **27 million images**, each paired with a long (~120 token) caption generated by **Qwen/Qwen2.5-VL-7B-Instruct**.

### Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="BLIP3o/BLIP3o-Pretrain-Long-Caption",
repo_type="dataset"
)
```

## Load Dataset without Extracting
You don’t need to unpack the .tar archives, use WebDataset support in 🤗datasets instead:

```
from datasets import load_dataset
import glob

data_files = glob.glob("/your/data/path/*.tar")
train_dataset = load_dataset(
"webdataset",
data_files=data_files,
cache_dir="/your/hf/cache",
split="train",
num_proc=128
)
```

Feel free to comment on it when you have any issues.

## BLIP3o 预训练长字幕数据集

该数据集集合包含**2700万张图像**，每张图像均搭配由**Qwen/Qwen2.5-VL-7B-Instruct**生成的长字幕（约含120个词元（Token））。

### 下载

python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="BLIP3o/BLIP3o-Pretrain-Long-Caption",
repo_type="dataset"
)

### 无需解压加载数据集

您无需解压.tar归档文件，只需使用🤗datasets中的WebDataset支持即可：

from datasets import load_dataset
import glob

data_files = glob.glob("/your/data/path/*.tar")
train_dataset = load_dataset(
"webdataset",
data_files=data_files,
cache_dir="/your/hf/cache",
split="train",
num_proc=128
)

若您遇到任何问题，欢迎留言反馈。