README.md

metadata

license: apache-2.0
task_categories:
  - image-to-text
  - text-to-image
tags:
  - monet
  - captions
  - synthetic-captions
  - qwen
  - multimodal
  - text-to-image-training
language:
  - en

MONET Captions Dataset

MONET (Massive, Open, Non-redundant and Enriched Text-to-image dataset) is a large-scale, curated image-text dataset designed for training text-to-image (T2I) systems.

• jasperai/monet

This repository provides an enriched, concise, and multi-VLM-synthesized caption dataset derived from MONET to improve training efficiency and text description quality.

🛠️ Dataset Construction

The dataset is built upon a sample of ~100,000 images sourced from agentlans/jasperai-monet-sample.

Processing Pipeline

1. Multi-VLM Caption Extraction: Individual captions were generated using a variety of vision-language models (e.g., Florence-2-Large, Gemini 2.5 Flash Lite, InternVL-3-8B, and ShareGPT4V-7B).
2. Qwen3.5-9B Merging: The individual model captions were aggregated and merged using Qwen3.5-9B to produce a single, highly detailed yet concise caption.

Merging Prompt

Merge the captions below into one short, accurate caption.

Rules:

* Include the key details from all captions.
* Remove repeated or unnecessary information.
* Keep the result clear, natural, and concise.
* Do not add or guess information.
* Output only the final caption.

Captions:
[CAPTIONS]

📂 Data Structure & Splits

The dataset is split into Train (80%) and Test (20%) sets.

Schema

• __key__: The unique ID. Matches the ID in the original dataset for retrieving the actual image.
• caption_<model_name>: Individual raw captions generated by respective VLMs.
• caption: The final synthesized, merged caption.
• abridged_caption: The caption shortened to roughly 75 CLIP tokens by selecting the highest-scoring tokens using TF-IDF.

Example Row

{
  "__key__": "055_cc61e2aebccd566d",
  "caption_florence-2-large": "The image shows a man with dreadlocks wearing a peach-colored t-shirt and sunglasses, smiling against a blue background.",
  "caption_gemini-2.5-flash-lite": "A close-up, profile shot shows a Black man from the chin up, wearing dark-lensed, brown-framed sunglasses. His mouth is open in a wide smile, revealing white teeth. He has dark, braided hair, with some braids pulled back. He wears a light peach-colored t-shirt and a thin, silver chain around his neck. The background is a solid, bright blue. The lighting is soft, highlighting the man's features and creating a natural, casual feel. The style is portraiture with a focus on the subject's expression and accessories.",
  "caption_internvl-3-8b": "The image shows a person with braided hair wearing dark sunglasses and a light-colored shirt. The background is a solid blue color, and the person is smiling. The overall mood of the image is cheerful and relaxed.",
  "caption_sharegpt4v-7b": "In the image, a young man is captured in a close-up shot against a vibrant blue background. His face is turned slightly to the left, and he is smiling, adding a warm and friendly vibe to the image. He is wearing a pair of sunglasses that are dark in color, possibly black or brown, and have a distinctive orange frame. The sunglasses are positioned on the top of his head, suggesting a casual and relaxed style. His hair is styled in braids, adding a unique touch to his overall appearance. He is wearing a pink",
  "caption": "A young man with dark, braided hair and bangs, wearing dark-lensed sunglasses with orange frames and a light peach-coloured t-shirt. He smiles warmly while his mouth is open, revealing white teeth. The background is a solid bright blue, and the lighting is soft, creating a cheerful, relaxed, and natural portraiture style.",
  "abridged_caption": "a young man with dark , braided hair and bangs , wearing dark - lensed sunglasses with orange frames and a light peach - coloured t - shirt . he smiles warmly while his mouth is open , revealing white teeth . the background is a solid bright blue , and the lighting is soft , creating a cheerful , relaxed , and natural portraiture style ."
}

⚠️ Limitations

• Dataset Bias: MONET's content leans predominantly toward safe-for-work (SFW) photos.
• VLM Limitations: Captions are dependent on the inherent text and vision recognition capabilities of the underlying models.
• Synthesis Biases: Potential stylistic or structural biases introduced during the Qwen3.5 text-merging process.
• Verification: Merged captions have not been exhaustively cross-checked manually against the original images.

📜 License

Distributed under the Apache 2.0 license, matching the original MONET dataset.