Vividbot/llava-pretrain-vi

---
license: mit
language:
- vi
size_categories:
- 100K<n<1M
---

# Structure
Each sample will have a structure as follows:
```
{
'id': Value(dtype='string', id=None),
'images': Value(dtype='binary', id=None),
'conversations': [{'from': Value(dtype='string', id=None), 'value': Value(dtype='string', id=None)}]
}

{
'id': '004309348',
'image': <image-bytes>,
'conversations': [{'from': 'human', 'value': 'Điều gì được minh họa trong hình ảnh này?\n<image>'}, {'from': 'gpt', 'value': 'bạn nghĩ có bao nhiêu sinh viên ở farbaut sử dụng sản phẩm thuốc lá'}]
}
```

# How To Use

## Convert binary objects
Because the returned video will be in bytes, here is a way to extract frames and fps:
```python
import io
import numpy as np
from PIL import Image
from datasets import load_dataset

def extract_image(image_bytes):
img = Image.open(io.BytesIO(image_bytes))
arr = np.asarray(img)
return arr

dataset = load_dataset("Vividbot/instruct500k_vi", name="all", streaming=True)

image_bytes = next(iter(dataset["train"]))["image"]
image = extract_image(image_bytes)
print(f"Image shape: {image.shape}")
```