license: apache-2.0
task_categories:
- image-to-text
- image-text-to-text
language: - en
- ja
- zh
tags: - art
- synthetic
- blip3o
- caption
- not-for-all-audiences
size_categories: - 1K<n<10K
Caption3o-Opt-Abliterated-5K
Caption3o-Opt-Abliterated-5K is a gated, synthetic image-caption dataset curated for training and evaluating multimodal (Image-to-Text / Image-Text-to-Text) models. It is a compact, filtered, and “abliterated” subset derived from Caption3o-Opt with an emphasis on experimental captioning strategies, covering diverse real-world and artistic content.
Dataset Summary
- Size: ~5,000 image–caption pairs
- Format: Parquet
- Image resolution: 512x512
- Languages: English, Japanese, Chinese
- Modalities: Image-to-Text, Image-Text-to-Text
- License: Apache-2.0
- Tags: Synthetic, Caption, Art, Abliterated, Not-For-All-Audiences
Each sample includes an image and a long-form, experimental caption intended to push the boundaries of caption quality, style, and interpretability. This “abliterated” version integrates noisy, synthetic, and stylistically diverse captions for robustness testing.
Features
image: 512x512 RGB imagecaption: Long-form text (multi-lingual: English / Japanese / Chinese)
Example:
A surreal rendering of a neon-lit alleyway where holographic koi fish swim above the pavement, reflecting off wet stone...
Use Cases
- Pretraining / finetuning multimodal models (e.g., BLIP, LLaVA, Qwen-VL, SigLIP)
- Evaluating model robustness on synthetic and “noisy” caption inputs
- Vision-language alignment research
- Creative visual storytelling and multilingual caption generation
How to Use
Load the dataset with Hugging Face datasets:
from datasets import load_dataset
dataset = load_dataset("strangerguardhf/Caption3o-Opt-Abliterated-5K", split="train")
Note: This is a gated dataset — you need access approval before use.
Citation
If you use this dataset, please cite the base dataset and this derivative:
BLIP3o/BLIP3o-Pretrain-Long-Caption
https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption
And reference this curated derivative:
Caption3o-Opt-Abliterated-5K by StrangerGuard