OpenCaption-FineGrained

# **OpenCaption-FineGrained** **OpenCaption-FineGrained** is a high-quality dense image captioning dataset containing fine-grained, long-form image descriptions synthesized using the **Qwen3.6 Multimodal** model. Rather than generating a single caption per image, the dataset follows a **multi-sample caption selection pipeline**, where multiple candidate captions are generated for every image and the highest-quality caption is selected using an automated quality filtering strategy. The resulting captions provide significantly richer visual understanding than conventional image descriptions, capturing scene composition, object relationships, spatial reasoning, attributes, actions, lighting, background context, and fine visual details. The dataset is built primarily from **publicly available images**, which constitute the majority of the input imagery, together with other publicly available datasets. # Features - Fine-grained, long-form image captions - Multiple caption candidates generated per image - Best caption selected through automated quality filtering - Rich descriptions covering objects, scenes, attributes, relationships, and context - Optimized for modern Vision-Language Models (VLMs) - Apache-2.0 licensed # Dataset Structure Each sample contains: | Column | Type | Description | |---------|------|-------------| | `image` | Image | Input image | | `response` | String | High-quality fine-grained caption | # Generation Pipeline Each image follows the workflow below: 1. Publicly available image collection 2. Multiple caption generations using **Qwen3.6 Multimodal** 3. Automated quality evaluation and ranking 4. Best caption retained This multi-generation strategy produces captions that are substantially more descriptive and consistent than single-pass caption generation. # Citation If you use this dataset in your research or projects, please cite: ```bibtex @misc{OpenCaptionFineGrained2026, author = {Prithiv Sakthi}, title = {OpenCaption-FineGrained}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/prithivMLmods}} } ``` **Dataset by** — https://huggingface.co/prithivMLmods