📊 MMLongBench-Doc Evaluation Results

Official evaluation results: GPT-4.1 (2025-04-14) & GPT-4o (2024-11-20) 📄 Paper: MMLongBench-Doc, NeurIPS 2024 Datasets and Benchmarks Track (Spotlight)

Accepted by NeurIPS 2024 Datasets and Benchmarks Track

We introduce the RePair puzzle-solving dataset, a large-scale real world dataset of fractured frescoes from the archaelogical campus of Pompeii. Our dataset consists of over 1000 fractured frescoes. The RePAIR stands as a realistic computational challenge for methods for 2D and 3D puzzle solving, and serves as a benchmark that enables the study of fractured object reassembly and presents new challenges for geometric shape understanding. Please visit our website for more dataset information, access to source code scripts and for an interactive gallery viewing of the dataset samples.

Semi Truths Dataset: A Large-Scale Dataset for Testing Robustness of AI-Generated Image Detectors (NeurIPS 2024 Track Datasets & Benchmarks Track)

Recent efforts have developed AI-generated image detectors claiming robustness against various augmentations, but their effectiveness remains unclear. Can these systems detect varying degrees of augmentation?

To address these questions, we introduce Semi-Truths, featuring 27, 600 real images, 223, 400 masks, and 1, 472, 700… See the full description on the dataset page: https://huggingface.co/datasets/semi-truths/Semi-Truths.

MmCows is a large-scale multimodal dataset for behavior monitoring, health management, and dietary management of dairy cattle.

The dataset consists of data from 16 dairy cows collected during a 14-day real-world deployment, divided into two modality groups. The primary group includes 3D UWB location, cows' neck IMMU acceleration, air pressure, cows' CBT, ankle acceleration, multi-view RGB images, indoor THI, outdoor weather, and milk yield. The secondary group contains measured UWB distances, cows' head direction, lying behavior, and health records.

MmCows also contains 20,000 isometric-view images from multiple camera views in one day that are annotated with cows' ID and their behavior as the ground truth. The annotated cow IDs from multi-views are used to derive their 3D body location ground truth.

Below is a portion of the whole dataset. More details of the dataset and benchmarks are available at https://github.com/neis-lab/mmcows.

This link offers faster and more reliable download: https://huggingface.co/datasets/neis-lab/mmcows

Brief overview video: https://www.youtube.com/watch?v=YBDvz-HoLWg

DOI: 10.57967/hf/5965 (cow)

BLEnD

This is the official repository of BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages (Submitted to NeurIPS 2024 Datasets and Benchmarks Track). 24/12/05: Updated translation errors25/05/02: Updated multiple choice questions file (v1.1)

  About

Large language models (LLMs) often lack culture-specific everyday knowledge, especially across diverse regions and non-English languages. Existing benchmarks for evaluating LLMs' cultural… See the full description on the dataset page: https://huggingface.co/datasets/nayeon212/BLEnD.

The paper has been accepted in NeurIPS 2024 (Dataset & Benchmark Track). paper repo

☠️ Warning: The samples presented by this paper may be considered offensive or vulgar.

❗️ Ethics Statement

The opinions and findings contained in the samples of our presented dataset should not be interpreted as representing the views expressed or implied by the authors. We acknowledge the risk of malicious actors attempting to reverse-engineer memes. We sincerely hope that users will employ the dataset responsibly and appropriately, avoiding misuse or abuse. We believe the benefits of our proposed resources outweigh the associated risks. All resources are intended solely for scientific research and are prohibited from commercial use.

📜 Chinese Harmful Meme

To adapt to the Chinese online environment, we introduce the definition of Chinese harmful memes:

Chinese harmful memes are multimodal units consisting of an image and Chinese inline text that have the potential to cause harm to an individual, an organization, a community, a social group, or society as a whole. These memes can range from offense or joking that perpetuate harmful stereotypes towards specific social entities, to memes that are more subtle and general but still have the potential to cause harm. It is important to note that Chinese harmful memes can be created and spread intentionally or unintentionally. They often reflect and reinforce underlying negative values and cultural attitudes on the Chinese Internet, which are detrimental from legal or moral perspectives.

📜 ToxiCN MM

According to the definition, we identify the most common harmful types of memes on Chinese platforms, including targeted harmful, general offense, sexual innuendo, and dispirited culture. We focus on these harmful types when constructing the dataset.

During the annotation, we label memes from two aspects: harmful types (i.e., the above four types) and modality combination (i.e., analyzing toxicity through fused or independent features, including Text-Image Fusion, Harmful Text, and Harmful Image). Finally, we present the ToxiCN MM dataset, which contains 12,000 samples.

Considering the potential risk of abuse, please fill out the following form to request the datasets: https://forms.gle/UN61ZNfTgMZKfMrv7. After we get your request, we will send the dataset to your email as soon as possible. The dataset labels and captions generated by GPT-4V have been saved as train_data_discription.json and test_data_discription.json in the ./data/ directory. Here we simply describe each fine-grain label.

📜 Detector

We present a Multimodal Knowledge Enhancement Detector for effective detection. It incorporates contextual information of meme content to enhance the detector's understanding of Chinese memes generated by the LLM. The requirements.txt file lists the specific dependencies of the project.

❗️ Licenses

This work is licensed under a Creative Commons Attribution- NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

Poster

https://github.com/user-attachments/assets/c3cb7793-33f2-4e3e-ad72-e0d84530c658" alt="poster_original">

Cite

If you want to use the resources, please cite the following paper. The camera-ready version of the paper will be released after the conference: ~~~ @article{lu2024towards, title={Towards Comprehensive Detection of Chinese Harmful Memes}, author={Lu, Junyu and Xu, Bo and Zhang, Xiaokun and Wang, Hongbo and Zhu, Haohao and Zhang, Dongyu and Yang, Liang and Lin, Hongfei}, journal={arXiv preprint arXiv:2410.02378}, year={2024} } ~~~

SciFIBench

  Jonathan Roberts, Kai Han, Neil Houlsby, and Samuel Albanie





  NeurIPS 2024

Note: This repo has been updated to add two splits ('General_Figure2Caption' and 'General_Caption2Figure') with an additional 1000 questions. The original version splits are preserved and have been renamed as follows: 'Figure2Caption' -> 'CS_Figure2Caption' and 'Caption2Figure' -> 'CS_Caption2Figure'.

  Dataset Summary

The SciFIBench (Scientific Figure… See the full description on the dataset page: https://huggingface.co/datasets/jonathan-roberts1/SciFIBench.

WikiDBs is an open-source corpus of 100,000 relational databases. We aim to support research on tabular representation learning on multi-table data. The corpus is based on Wikidata and aims to follow certain characteristics of real-world databases.

WikiDBs was published as a spotlight paper at the Dataset & Benchmarks track at NeurIPS 2024.

WikiDBs contains the database schemas, as well as table contents. The database tables are provided as CSV files, and each database schema as JSON. The 100,000 databases are available in five splits, containing 20k databases each. In total, around 165 GB of disk space are needed for the full corpus. We also provide a script to convert the databases into SQLite.

🌿 Welcome to the GeoPlant dataset hub on Kaggle! 🌍

This dataset provides pre-extracted features from multimodal environmental data and expert-verified species observations, ready to be integrated into your models. Whether you're here for research, experimentation, or competition, you're in the right place!

🔎 Check out the key resources below to get started: | Resource | Description | Link | | ------------------------------ | -------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | 📄 Dataset Paper | NeurIPS 2024 paper detailing the dataset, benchmark setup, etc. | NeurIPS Paper (PDF) | | | 🧠 GitHub Repository | Codebase with data loaders, baseline models, and utilities | GeoPlant Repo | | 🚀 Starter Notebooks | Baseline models, multimodal pipelines, and training scripts | GeoPlant Code on Kaggle | | 📦 Full Dataset | All provided data including the Presence-Only (PO) species observations. | GeoPlant Seafile |

Observations data

The species related training data comprises: 1. Presence-Absence (PA) surveys: including around 90 thousand surveys with roughly 10,000 species of the European flora. The presence-absence data (PA) is provided to compensate for the problem of false-absences of PO data and calibrate models to avoid associated biases. 2. Presence-Only (PO) occurrences: combines around five million observations from numerous datasets gathered from the Global Biodiversity Information Facility (GBIF, www.gbif.org). This data constitutes the larger piece of the training data and covers all countries of our study area, but it has been sampled opportunistically (without standardized sampling protocol), leading to various sampling biases. The local absence of a species among PO data doesn't mean it is truly absent. An observer might not have reported it because it was difficult to "see" it at this time of the year, to identify it as not a monitoring target, or just unattractive.

There are two CSVs with species occurrence data on the Seafile available for training. The detailed description is provided again on SeaFile in separate ReadME files in relevant folders. - The PO metadata are available in PresenceOnlyOccurences/GLC24_PO_metadata_train.csv. - The PA metadata are available in PresenceAbsenceSurveys/GLC24_PA_metadata_train.csv.

https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1518097%2Fcf0b0ee7f4ab8c1f7944fd7b3cd89d81%2FDataComposition.png?generation=1718369587083645&alt=media" alt="">

Environmental data

Besides species data, we provide spatialized geographic and environmental data as additional input variables (see Figure 1). More precisely, For each species observation location, we provide: 1. Satellite image patches: 3-band (RGB) and 1-band (NIR) 128x128 images at 10m resolution. 2. Satellite time series: Up to 20 years of values for six satellite bands (R, G, B, NIR, SWIR1, and SWIR2). 3. Environmental rasters Various climatic, pedologic, land use, and human footprint variables at the European scale. We provide scalar values, time-series, and original rasters from which you may extract local 2D images.

There are three separate folders with the relevant data on the Seafile available for training. The detailed description is provided below and again on SeaFile in separate "Readme" files in relevant folders. - The Satellite image patches in ./SatellitePatches/. - The Satellite time series in ./SatelliteTimeSeries/. - The Environmental rasters in ./EnvironmentalRasters/.

Figure. Illustration of of the environmental data for an occurrence (glcID=4859165) collected in northern Switzerland (lon=8.5744;lat=47.7704) in 2021. A. The 1280x1280m satellite image patches were sampled in 2021 around the observation. B. Quarterly time series of six satellite ...

MatSeg Dataset and benchmark for zero-shot material state segmentation.

MatSeg Benchmark containing 1220 real-world images and their annotations is available at MatSeg_Benchmark.zip the file contains documentation and Python readers.

MatSeg dataset containing synthetic images with infused natural images patterns is available at MatSeg3D_part_*.zip and MatSeg3D_part_*.zip (* stand for number).

MatSeg3D_part_*.zip: contain synthethc 3D scenes

MatSeg2D_part_*.zip: contain syntethc 2D scenes

Readers and documentation for the synthetic data are available at: Dataset_Documentation_And_Readers.zip

Readers and documentation for the real-images benchmark are available at: MatSeg_Benchmark.zip

The Code used to generate the MatSeg Dataset is available at: https://zenodo.org/records/11401072

Additional permanent sources for downloading the dataset and metadata: 1, 2

Evaluation scripts for the Benchmark are now available at:

https://zenodo.org/records/13402003 and https://e.pcloud.link/publink/show?code=XZsP8PZbT7AJzG98tV1gnVoEsxKRbBl8awX

License

This dataset, including all its components, is released under the CC0 1.0 Universal (CC0 1.0) Public Domain Dedication. To the extent possible under law, the authors have dedicated all copyright and related and neighboring rights to this dataset to the public domain worldwide. This dedication applies to the dataset and all derivative works.

The MatSeg 2D and 3D synthetic were generated using the open-images dataset which is licensed under the https://www.apache.org/licenses/LICENSE-2.0. For these components, you must comply with the terms of the Apache License. In addition, the MatSege3D dataset uses Shapenet 3D assets with GNU license.

Example Usage:

An Example of a training and evaluation code for a net trained on the dataset and evaluated on the benchmark is given at these urls: 1, 2

This include an evaluation script on the MatSeg benchmark.

Training script using the MatSeg dataset.

And weights of a trained model

Paper:

More detail on the work ca be found in the paper "Infusing Synthetic Data with Real-World Patterns for
Zero-Shot Material State Segmentation"

Croissant metadata and additional sources for downloading the dataset are available at 1,2

[NeurIPS D&B 2024 Spotlight] ChronoMagic-Bench: A Benchmark for Metamorphic Evaluation of Text-to-Time-lapse Video Generation

If you like our project, please give us a star ⭐ on GitHub for the latest update.

  💡 Description

Repository: Code, Page, Data Paper: https://huggingface.co/papers/2406.18522 Point of Contact: Shenghai Yuan

  ✏️ Citation

If you find our paper and code useful in your research, please consider giving a star and citation.… See the full description on the dataset page: https://huggingface.co/datasets/BestWishYsh/ChronoMagic-Bench.

SentinelKilnDB - A Large-Scale Dataset and Benchmark for OBB Brick Kiln Detection in South Asia Using Satellite Imagery

NeurIPS 2025 Datasets & Benchmarks Track

Abstract

Air pollution was responsible for 2.6 million deaths across South Asia in 2021 alone, with brick manufacturing contributing significantly to this burden. In particular, the Indo-Gangetic Plain; a densely populated and highly polluted region spanning northern India, Pakistan, Bangladesh, and parts of Afghanistan sees brick kilns contributing 8–14% of ambient air pollution. Traditional monitoring approaches, such as field surveys and manual annotation using tools like Google Earth Pro, are time and labor-intensive. Prior ML-based efforts for automated detection have relied on costly high-resolution commercial imagery and non-public datasets, limiting reproducibility and scalability. In this work, we introduce SENTINELKILNDB, a publicly available, hand-validated benchmark of 62,671 brick kilns spanning three kiln types Fixed Chimney Bull’s Trench Kiln (FCBK), Circular FCBK (CFCBK), and Zigzag kilns—annotated with oriented bounding boxes (OBBs) across 2.8 million km2 using free and globally accessible Sentinel-2 imagery. We benchmark state-of-the-art oriented object detection models and evaluate generalization across in-region, out-of-region, and super-resolution settings. SENTINELKILNDB enables rigorous evaluation of geospatial generalization and robustness for low-resolution object detection, and provides a new testbed for ML models addressing real-world environmental and remote sensing challenges at a continental scale. Datasets and code are available in SentinelKiln Dataset and SentinelKiln Benchmark, under the Creative Commons Attribution–NonCommercial 4.0 International License.

https://storage.googleapis.com/kagglesdsdata/datasets/8335452/13157427/statistics.png?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=databundle-worker-v2%40kaggle-161607.iam.gserviceaccount.com%2F20251021%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20251021T155911Z&X-Goog-Expires=345600&X-Goog-SignedHeaders=host&X-Goog-Signature=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" alt="Statistics">

Useful Links

Project Page - https://lnkd.in/dn2SKwWv
Official Paper - https://neurips.cc/virtual/2025/poster/121530
Github - https://github.com/rishabh-mondal/NeurIPS_2025
Sustainability Lab - https://sustainability-lab.github.io

For questions or collaborations, please contact:

Rishabh Mondal - rishabh.mondal@iitgn.ac.in
Nipun Batra - nipun.batra@iitgn.ac.in

Dataset Overview

This dataset contains Sentinel-2 satellite imagery focused on identifying and classifying brick kilns across the Indo-Gangetic Plain and neighboring South Asian countries, including Afghanistan, Pakistan, and Bangladesh.

• Imagery Source: Sentinel-2 (Surface Reflectance)
• Image Size: 128 × 128 pixels
• Spatial Resolution: 10 m/pixel
• Timeframe: November 2023 – February 2024
• Geographic Coverage: Indo-Gangetic Plain, Afghanistan, Pakistan, Bangladesh
• Overlap: 30-pixel overlap between patches
• File Naming Convention: lat,lon.png and lat,lon.txt

Classes

• CFCBK – Continuous Fixed Chimney Bull’s Trench Kiln
• FCBK – Fixed Chimney Bull’s Trench Kiln
• Zigzag – Zigzag Kiln

Annotation Formats

• YOLO OBB:
  class_name, x1, y1, x2, y2, x3, y3, x4, y4

• YOLO AA:
  class_name, x_center, y_center, width, height

• DOTA Format:
  x1, y1, x2, y2, x3, y3, x4, y4, class_name, difficult

Dataset Splits

The dataset is split using a class-wise stratified approach for balanced representation.

Each split contains separate folders for images and annotations:

dataset/
├── train/
│  ├── images/
│  └── labels/
├── val/
│  ├── images/
│  └── labels/
└── test/
  ├── images/
  └── labels/

...

The dataset contains UAV footage of wild antelopes (blackbucks) in grassland habitats. It can be mainly used for two tasks: Multi-object tracking (MOT) and Re-Identification (Re-ID). We provide annotations for the position of animals in each frame, allowing us to offer very long videos (up to 3 min) completely annotated while maintaining the identity of each animal in the video. The Re-ID dataset offers two videos, that capture the movement of some animals simultaneously from two different UAVs. The Re-ID task is to find the same individual in two videos taken simultaneously from a slightly different perspective. The relevant paper will be published in the NeurIPS 2024 Dataset and Benchmarking Track. https://nips.cc/virtual/2024/poster/97563 Resolution: 5.4 K MOT: 12 videos ( MOT17 Format) Re-ID: 6 sets (each with a pair of drones) (Custom) Detection: 320 Images (COCO, YOLO)

ChronoMagic Dataset

This dataset contains time-lapse video-text pairs curated for metamorphic video generation. It was presented in the paper ChronoMagic-Bench: A Benchmark for Metamorphic Evaluation of Text-to-Time-lapse Video Generation. Project page: https://pku-yuangroup.github.io/ChronoMagic-Bench

  Usage

cat ChronoMagic-ProH_part_* > ChronoMagic-ProH.zip unzip ChronoMagic-ProH.zip

[NeurIPS D&B 2024 Spotlight] ChronoMagic-Bench: A Benchmark for Metamorphic… See the full description on the dataset page: https://huggingface.co/datasets/BestWishYsh/ChronoMagic-ProH.

CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs

NeurIPS 2024 🏠Home (🚧Still in construction) | 🤗Data | 🥇Leaderboard | 🖥️Code | 📄Paper This repo contains the full dataset for our paper CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs, which is a diverse and challenging chart understanding benchmark fully curated by human experts. It includes 2,323 high-resolution charts manually sourced from arXiv preprints. Each chart is… See the full description on the dataset page: https://huggingface.co/datasets/princeton-nlp/CharXiv.