datasets

Mundo AI secures $20M in Series A funding to build datasets, evaluations, and research for AI perception across audio, video, and emerging modalities.

Introduces a list on GitHub named Awesome Public Datasets, which compiles high-quality public datasets from around the world, helping data analysts save time searching for data.

A new open-source repository curates a list of egocentric (first-person vision) datasets, updated to include recent releases from 2023–2026 with sections on VLA papers and broken links removed.

This paper reviews multi-turn conversational AI across text, audio, and multimodal systems, analyzing datasets, modeling paradigms, and evaluation methods, while highlighting key challenges such as persistent memory and cross-turn grounding.

HKUST and Meituan jointly release the open-source poster generation model PosterCraft, optimizing Chinese text rendering and layout via multi-stage training, and offering complete datasets and code.

A new tool for browsing over 10,000 neuroscience datasets has been released, featuring nearly 2,000 datasets from Hugging Face.

Hugging Face's mid-2026 report analyzes the open-model ecosystem, noting that Chinese labs now lead in releasing the largest open models, Qwen has become the community's base model, and small models plus agents are driving practical adoption.

The article points out that real robot interaction data is the key bottleneck for VLA deployment. Model architectures are converging, but data acquisition is difficult. A few companies own proprietary data that forms a moat, while open-source datasets such as Open X-Embodiment and DROID are available for reference and validation.

Argues that the most valuable robotics acquisition target is the company owning the best real-world manipulation dataset, as models converge but data remains scarce and concentrated among a few players.

NVIDIA released the NeMo Gym conversational tool-use assets on Hugging Face, including golden policy/tool reference pairs and prompt histories for the pipeline.

A GitHub repository aggregates dataset download sites for large model and generative AI training, covering text-to-text, text-to-image, and text-to-video, categorized by region and type, with both free and paid resources.

Sergey Levine highlights the importance of learning from suboptimal robot data and promotes OopsieData, a community effort to collect and open-source the abundant but underused failed robot interaction clips.

A developer shares the challenge of creating a gold standard evaluation dataset for an AI product with no users, considering synthetic data generation and adversarial testing to avoid post-launch restructuring.

This paper investigates when machine learning outperforms traditional value sorting for exposure-weighted shipment prioritization using three datasets.

This survey examines computational humor understanding in multimodal LLMs, covering methods, datasets, evaluation protocols, and challenges such as shortcut-prone evaluation and weak evidence grounding.

This position paper argues that ground truth datasets in machine learning are not objective truths but human constructions shaped by choices, and advocates for articulating these choices to improve reliability, transparency, and accountability.

This project open-sources datasets that can be used for algorithm training and research.

The Well is an open-source collection of 15TB of physics simulation datasets spanning 16 domains, created by Flatiron Institute and 11 universities, designed to train PDE surrogate models, enabling researchers to replace expensive supercomputer simulations with neural networks.

A systematic review of non-social media free-text datasets for mental health disorder detection, identifying biases and gaps in current resources.

LeRobot v0.6.0 is a major release of Hugging Face's robot learning library, adding world model policies (VLA-JEPA, FastWAM, LingBot-VA), new VLAs, reward models, six simulation benchmarks, depth sensing, VLM-powered dataset annotation, custom video encoding, cloud training, and a deployment CLI for human-in-the-loop corrections.