datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
LLaVA-ReCap-CC12MRecap-DataComp-1B
Dataset Card for Recap-DataComp-1B
Recap-DataComp-1B is a large-scale image-text dataset that has been recaptioned using an advanced LLaVA-1.5-LLaMA3-8B model to enhance the alignment and detail of textual descriptions.
Dataset Details
Dataset Description
Our paper aims to bridge this community effort, leveraging the powerful and open-sourced LLaMA-3, a GPT-4 level LLM.
Our recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered… See the full description on the dataset page: https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B.VoRA-Recap-GLDv2-1.4MImagenet21K_RecaptionThis dataset is the entire 21K ImageNet dataset with about 13 million examples and about 19 thousand classes as strings
(for some reason it only had ~19K classes instead of 21K).
If you want an even larger set of images, I have a recaptioned CC12M and ImageNet dataset: https://huggingface.co/datasets/gmongaras/CC12M_and_Imagenet21K_Recap
The images are in PNG format. They can be decoded like in the following example
import io
from PIL import Image
Image.open(io.BytesIO(row["image"]))
where… See the full description on the dataset page: https://huggingface.co/datasets/gmongaras/Imagenet21K_Recaption.midjourney-v6-recap
Midjourney v6 Recaptioned
~1.2M Midjourney v6 images with captions from three VLMs:
llava: Original LLaVA captions from the source dataset
gemini: Gemini Flash 1.5 captions
qwen3: Qwen3 VL 8B captions
Caption coverage
llava: available for all 1,235,432 images (from original dataset)
gemini and qwen3: available for 1,017,105 images (82.3%)
Source
Based on brivangl/midjourney-v6-llava.
cc12m-recaptionedLLaVA-ReCap-CC3Mcc12m-wds-coco-recaptioned
CC12M WebDataset with COCO-style Recaptions
A large-scale image-text dataset containing 3 million images from Conceptual Captions 12M (CC12M) with COCO-style factual descriptions generated using NVIDIA Nemotron Nano 12B v2 VL.
Dataset Overview
Base Dataset: pixparse/cc12m-wds - Conceptual Captions 12M (CC12M)
Images: 3,000,000+ high-quality internet images
Recaption Model: NVIDIA Nemotron Nano 12B v2 VL
Recaption Style: COCO-style factual descriptions (20 words average)… See the full description on the dataset page: https://huggingface.co/datasets/undefined443/cc12m-wds-coco-recaptioned.CC12M_and_Imagenet21K_RecapI removed all low quality data and uploaded it here
This dataset is the entire 21K ImageNet dataset with about 13 million examples and about 19 thousand classes as strings
(for some reason it only had ~19K classes instead of 21K) as well as the entire CC12M dataset, recaptioned.
If you just want the recaptioned Imagenet dataset, I have that here
I obtained the CC12M form others. CC12M is a dataset with 12 million images created in 2021. Unfortunately the downloader provided by Google has many… See the full description on the dataset page: https://huggingface.co/datasets/gmongaras/CC12M_and_Imagenet21K_Recap.recap-datacomp-12m-wdsStable_Diffusion_3_RecaptionThis dataset is the one specified in the stable diffusion 3 paper which is composed of the ImageNet dataset and the CC12M dataset.
I used the ImageNet 2012 train/val data and captioned it as specified in the paper: "a photo of a 〈class name〉" (note all ids are 999,999,999)
CC12M is a dataset with 12 million images created in 2021. Unfortunately the downloader provided by Google has many broken links and the download takes forever.
However, some people in the community publicized the dataset.… See the full description on the dataset page: https://huggingface.co/datasets/gmongaras/Stable_Diffusion_3_Recaption.recap-t2i-evaluation-sample-2026
Recaptioned T2I Supervision Evaluation Sample
This repository is the small reviewer-inspection companion to the full anonymous caption-metadata release. The full release is hosted separately at https://huggingface.co/datasets/Anonymous1477/recap-t2i-evaluation-metadata-2026; this repository stays under the large-dataset sample threshold and gives reviewers a direct way to inspect redacted caption metadata, join structure, and selected image-conditioned audit packages.… See the full description on the dataset page: https://huggingface.co/datasets/Anonymous1477/recap-t2i-evaluation-sample-2026.Recap-OmniCAD
Recap-OmniCAD
Recap-OmniCAD is a large-scale CAD dataset introduced in the CVPR 2026 paper:
Pointer-CAD: Unifying B-Rep and Command Sequences via Pointer-based Edges & Faces SelectionDacheng Qi, Chenyu Wang, Jingwei Xu, Tianzhe Chu, Zibo Zhao, Wen Liu, Wenrui Ding, Yi Ma, Shenghua Gao
Data Extraction
If you have downloaded a set of tar.gz archives under ./data, you can extract each archive into a folder with the same name as the archive file.
For example… See the full description on the dataset page: https://huggingface.co/datasets/Snitro/Recap-OmniCAD.LLaVA-ReCap-558KPenguin-Recap-I
Penguin-Recap-I
Penguin-Recap-I publishes recap metadata only. The repository does not contain
image binaries.
Included subsets
subset
collection
local source roots
expected records
datacomp_coyo_penguin
DataComp + COYO Penguin recap
datamultimodal/IMAGE/datacomp_1b, datamultimodal/IMAGE/coyo_700m
57,618,155
sa1b_penguin
SA-1B Penguin recap
datamultimodal/IMAGE/SA-1B
9,254,501
openimages_penguin
OpenImages Penguin recap… See the full description on the dataset page: https://huggingface.co/datasets/tencent/Penguin-Recap-I.qom_recapsdfsdf
wikiart_recaptionWikiArt Dataset captioned using vikhyatk/moondream2 model with prompt : Generate a short, simple and only visually descriptive caption for this image.
IN1k256-AR-buckets-bfl16latents_dc-ae-f32c32-sana-1.0_recapVoRA-Recap-8MLLaVA-ReCap-676KThis is an integrated version of LLaVA-ReCap, sourced from lmms-lab/LLaVA-ReCap-558K and lmms-lab/LLaVA-ReCap-118K.
In this version, the conversations field has been split into two separate fields: prompt and response. Additionally, the <image> special token has been removed to facilitate customization.
Inspired by the original paper, the prompt field has been further expanded with human-crafted variations. Specifically, each prompt is sampled from one of the following 30 instructions:… See the full description on the dataset page: https://huggingface.co/datasets/LimeryJorge/LLaVA-ReCap-676K.datacomp_recap_metadata2Recap-Datacomp-1B_tars_part7Final size: 7,236,721, samples per tar: 10000
cc3m-recap-wdsRecap-DataComp-1B_split_3Penguin-Recap-V
Penguin-Recap-V
Penguin-Recap-V provides Multi-granularity video annotation. This figure illustrates the alignment between visual content and textual descriptions across three temporal scales: Dense time-level, Paragraph-level, and Video-level.
Included subsets
subset
source collection
videos / clips
expected rows
source jsonl
sharegpt4video
ShareGPT4Video
40,145
120,435
sharegpt4video/predictions_process_relative.jsonl
shortvideo
ShortVideo
147,326
441,978… See the full description on the dataset page: https://huggingface.co/datasets/tencent/Penguin-Recap-V.wit_1m_recaptioned
Back into Plato’s Cave: Examining Cross-modal Representational Convergence at Scale
Image–text dataset derived from Wikipedia-based Image Text (WIT) with original and Gemini-generated captions, introduced in Back into Plato’s Cave: Examining Cross-modal Representational Convergence at Scale.
Configs
wit_1024
A fixed set of 1,024 query samples used for alignment evaluation.
from datasets import load_dataset
ds = load_dataset("askoepke/wit_1m_recaptioned"… See the full description on the dataset page: https://huggingface.co/datasets/askoepke/wit_1m_recaptioned.Recap-DataComp-1B_split_4Recap-DataComp-1B-FoodOrDrink
Recap-DataComp-1B: Food or Drink
A filtered subset of Recap-DataComp-1B containing 106,230,157 rows classified as food/drink content, enriched with structured food/drink extraction from FoodExtract-v2.
Overview
Count
Percentage
Total rows
106,230,157
100%
Food/drink (Stage 5 label)
96,618,895
91.0%
Not food/drink (Stage 5 label)
9,611,262
9.0%
FoodExtract (re_caption): food/drink
79,519,489
74.9%
FoodExtract (re_caption): not food/drink
26,710,156… See the full description on the dataset page: https://huggingface.co/datasets/mrdbourke/Recap-DataComp-1B-FoodOrDrink.VoRA-Recap-29MCOCONut-PanCap-Recaptioned
COCONut-PanCap Re-captioned (train2017)
Regenerated captions for the train2017 split of COCONut-PanCap
(paper,
dataset).
The released captions contain phrases bound to the wrong segment, references to objects that
are not in the image, redundant restatements, and untagged summary sentences that carry no
grounding. We regenerated every caption and repaired the remaining errors, keeping the file
layout of the original release.
Only the caption text changed. Mask ids… See the full description on the dataset page: https://huggingface.co/datasets/Panorama-grounding/COCONut-PanCap-Recaptioned.