Open Dataset Pulse

New Biological Datasets

• 🔥 MetaOmics-10T: MetaOmics-10T is a foundational dataset with 25K diverse microbiome samples and 10M reads, passed...

Key trend in massive omics data for biology AI:

• Trillion Gene Atlas by Basecamp Research (w/ Anthropic, NVIDIA): trillion-gene scale from 100M+...

SocialOmni introduces a benchmark for audio-visual social interactivity in omni models, advancing evaluation of multimodal capabilities in social scenarios. Join the paper discussion for latest insights.

Massive public dataset releases turbocharge genomics AI pipelines:

• AlphaFold expansion: Millions of protein complexes added to the database,...

Rahul Sarkar's 32-minute workshop talk explores datasets and challenges in accurate graph inference from images and documents—key for multimodal document understanding research. Watch on YouTube via ICMS channel.

Rising trend in domain-specific LLM evals:

• PeruMedQA draws from real Peruvian physician specialty exams for benchmarking and fine-tuning
• SciZoom...

Combine foundation-model surrogates for active learning with full pipeline best practices:

• Skip scratch training: Leverage meta-learned priors from...

FHIBE debuts as the first publicly available, consent-driven, globally diverse dataset for bias evaluation in human-centric CV tasks.

Key...

Streamlines high-dimensional bio analysis by training ML models like GateNet on labeled data for reproducible gating.

Key features:

• Versioned...

New Benchmarks and Datasets

• 🔥 MMOU: MMOU is a massive multi-task omni understanding and reasoning benchmark for long and complex real-world...

Emerging open tools drive scalable synthetic dataset trend, skipping real-world collection:

• OpenSeeker uses fact-grounded scalable controllable QA...

Exciting new dataset for spike-based visual recognition:

• NE15-MNIST is a spiking version of the original non-spiking MNIST handwritten digits...

Autonomous trucking demands enormous datasets for robust AI:

• A 20-minute test run yields ~100TB of raw sensor data from cameras, lidar, radar, GPS,...

Fresh benchmark for small AI models just dropped on Hugging Face:

• 18 models from 12 makers ranked
• Scores via HF Inference API; speed tested on 5 prompts × 3 rounds
• Released March 2026—track for compact model gains

Key unlock for AI researchers: Purdue’s Data to Science Initiative (D2S) joins AWS Open Data Sponsorship Program, making global geospatial UAV...

Key breakthroughs in ag AI datasets:

• $7M seed funding emerges Living Models from stealth to build foundation models shortening crop breeding...

Rising trend in multimodal evaluation datasets tackling complex scenarios:

• CoMa-20K: Benchmark for contextual massing generation with VLMs, pairing...

Best practices shift: Quality curation over sheer volume for research-grade models.

• Gold datasets enable strong models via disciplined data...

Hands-on demo for synthetic data to fix gaps:

• Datasets lack edge cases like school buses on highways.
• Use Gemini image editing: Prompt to add...

OpenFold announces a major update and public release of training data. The OpenFold3 white paper reports competitive performance versus AlphaFold3 across broad modalities and tasks – ideal for open protein folding research.