Sample page. This Hugging Face repository hosts only a 10-example-per-subset sample of RefinedVision, for quick browsing on the Hub. The full dataset is hosted on GitLab: https://gitlab.llm-jp.nii.ac.jp/datasets/refinedvision

RefinedVision is a refined version of HuggingFaceM4/FineVision that removes subsets with licensing or quality issues and regenerates QA pairs for low-quality subsets. As a result, RefinedVision contains 124 of FineVision's 185 subsets (66 as-is and 58 regenerated), totaling 11,965,359 examples.

This dataset is used to train llm-jp/llm-jp-4-vl-9b.

Construction Pipeline

FineVision by HuggingFace is a multimodal SFT dataset built from 185 subsets, but it contained several issues:

• Examples with license / terms-of-use problems (e.g., OpenAI TOU)
• Low image quality (e.g., synthetic chart/diagram images whose overlapping labels are unreadable)
• Low QA quality

For each FineVision subset, we manually inspected its license and checked its image / QA quality by randomly sampling about 5 examples and inspecting them visually. Based on this, we split the subsets into three groups:

RefinedVision removes the 61 problematic subsets. Of the remaining 124 subsets, the QA of 58 was regenerated with Qwen3.5 397B-A17B using a per-category prompt (the original images are kept), while the other 66 are used as-is. The exact prompts, per-category policy, and sampling configuration are in regeneration.md.

FineVision contains no reasoning traces. For the regenerated Math / Science / Chart & Table subsets, we additionally attach a thinking part (a summarized reasoning trace).

Usage

The full dataset is hosted on GitLab, so first clone it locally:

git lfs install
git clone https://gitlab.llm-jp.nii.ac.jp/datasets/refinedvision.git

from datasets import load_dataset

# one subset
ds = load_dataset("./refinedvision", "chartqa", split="train", streaming=True)
example = next(iter(ds))
print(example)
# {
#     "id": "0",                             # example id within the subset
#     "source": "chartqa",                   # subset name
#     "generator": "",                       # QA generator model ("Qwen/Qwen3.5-397B-A17B" if regenerated, "" otherwise)
#     "regenerated": False,                  # whether the QA was regenerated
#     "images": [<PIL.Image.Image>],         # list of images
#     "conversations": [                     # multi-turn QA
#         {"from": "human", "value": "...", "think": ""},
#         {"from": "gpt", "value": "...", "think": "..."},  # "think" is non-empty only for regenerated Math / Science / Chart & Table subsets
#     ],
# }

Per-category statistics: number of examples in FineVision (185 subsets) vs RefinedVision (124 subsets)

Per-subset statistics (124 subsets)

The licenses were verified manually on a best-effort basis and may contain errors.

License

The license of each component of RefinedVision is as follows:

This dataset is provided for use within the scope permitted under Article 30-4 of the Japanese Copyright Act.

Core contributors

Issa Sugiura, Keisuke Nakao, Koichi Akabe, Zhengdong Yang, Takashi Kodama