Ethlake/llava-665k

LLaVA-v1.5 Mix665K — Arrow (images embedded)

The LLaVA-v1.5 visual instruction-tuning mixture (llava_v1_5_mix665k) converted to a 🤗 datasets Arrow dataset with image bytes embedded. 665,298 examples (624,610 image–text + 40,688 text-only).

⚠️ This repo is a raw save_to_disk Arrow snapshot. The dataset viewer and `load_dataset()` do not work here — load it with `load_from_disk` as shown below.

Loading

from huggingface_hub import snapshot_download
from datasets import load_from_disk

path = snapshot_download("Ethlake/llava-665k", repo_type="dataset")
ds = load_from_disk(path)
print(ds)        # 665298 rows
print(ds[0])

Schema

Example row:

{
 'id': '000000033471',
 'subset': 'coco',
 'images': <PIL.Image RGB 640x480>,
 'texts': [
   {'user': 'What are the colors of the bus in the image?',
    'assistant': 'The bus in the image is white and red.'},
   ...
 ],
}

Subset distribution

Conversion notes

• —Image bytes are embedded directly in the Arrow shards — fully self-contained, no separate image folders required.
• —355 OCR-VQA images referenced as .jpg but stored on disk as .gif/.png were re-encoded to JPEG (RGB) and embedded.
• —The original flat conversations ({from, value}) were paired into {user, assistant} turns, and the <image> placeholder token was stripped (chat templates re-insert it at train time).
• —Text-only samples have images = None.

Source

Derived from the LLaVA-v1.5 llava_v1_5_mix665k instruction mixture. Underlying images come from COCO, Visual Genome, GQA, OCR-VQA and TextVQA; language-only data from ShareGPT. Please respect the original licenses of each upstream source.