Tag
Nvidia 发布了免费开源权重语音转写模型 Nemotron 3.5 Transcribe（30B 参数，支持 40 种语言）以及 Granary 开源语音数据集 v2，瞄准会议记录和医疗转写等企业应用场景。该模型声称在 Open ASR 等基准上超越竞争对手，且可在单芯片上运行。
TextRIRO 提出仅使用真实数据进行场景文本编辑的方法，通过图像扰动（Shuffle&Crop）构建自配对训练数据并结合条件扩散模型，显著缩小合成到真实的域差距，并发布了 TextRIRO-3M 大规模真实数据集用于长文本识别。
3DCodeVerse introduces a research ecosystem with a 2M-scale dataset of executable 3D code, a multi-agent DAG harness for automated generation and verification of 3D content, and empirical guidance showing fine-tuned open models achieve a 90% win rate against Claude Design (3D Object).
Maniformer 的新论文提出 MEgoVista，一套基于头戴式装置的物理 AI 数据系统，通过立体视觉和全序列优化，将手部手指误差降至 4.29 毫米、头部跟踪误差控制在 0.65–2.83 毫米，并经 26 相机动作捕捉实验室验证，为机器人学习提供可测量的人类动作数据。
Show HN: carrier-explode 2.0 decodes and compares carrier settings (APNs, VoLTE, Wi-Fi Calling, 5G) from iPhone, Pixel and Galaxy firmware, with an API and CC0 daily datasets available. The project is open source and seeking decoder contributions.
OpenAI reportedly plans to release hundreds of AI-solved math problems, drawing criticism from mathematicians who argue such outputs devalue human mathematical reasoning and proper proof standards.
Xiaomi open-sourced ~7K RL environments used to train MiMo on HuggingFace, while billxbf uploaded a 7,780-entry MiMo RL dataset and released Skill2Env with 7,496 additional RL datasets to Harbor Hub — 15,276 RL data points made free that would otherwise cost millions from vendors.
OneStreamer introduces a unified framework for streaming video LLMs that jointly learns query-independent evidence recording and task response via Proactive Hierarchical Caption Memory and Proactive State Transition Learning, delivering a 4B model that achieves state-of-the-art results across eight streaming video understanding benchmarks along with the OneStreamer-1M dataset.
4Director introduces a video world model conditioned on an explicit 4D scene representation, where each object is reconstructed as a canonical mesh and driven by per-frame rigid 3D transformations, combined with a MotionAdapter that renders the geometric scaffold into view-consistent video. The authors also release the RealCOD-Rigid dataset and the IG-IoU metric, showing improved camera and object control over prior methods.
Hugging Face Hub now features a dedicated filter for RL Environments, letting datasets tagged as environments be discovered, versioned, and run across frameworks like Harbor, Verifiers, and NVIDIA NeMo Gym without new registries or repo types.
TinyFish launched a 15-company alliance to provide licensed datasets for AI agents, highlighting data as a competitive moat in AI development.
The article discusses the process of utilizing open training data from Hugging Face to train Marin’s 535B model, which involved 25T tokens from 152 datasets with permissible licenses.
This paper proposes a training recipe for tabular foundation models using a Joint-Embedding Predictive Architecture (JEPA) to address latent term collapse, and compares its performance to a value-only approach across 147 real datasets.
The article proposes that type diversity in datasets explains why Transformers struggle with structural compositional generalisation, and validates this by varying diversity in COGS and SLOG datasets.
The article introduces AMDKernelVault, an open dataset and training framework for AMD GPU kernel optimization, featuring large-scale HIP and Triton kernels and agent-driven pipelines for generating and validating kernels.
The paper introduces E-CONAN benchmarks for Arabic textual entailment and natural inference, featuring diverse datasets and evaluations of multilingual and LLM models to enrich research in the domain.
OpenAI announces that premium financial data is now included, featuring datasets from Daloopa, PitchBook, and LSEG News.
This article is a detailed tutorial on the Hugging Face platform, guiding you from beginner to advanced, explaining how to find and use open-source models, datasets, and applications.
Hugging Face is exploring a sale valued at around $13 billion, raising concerns about potential policy changes for open models due to third-party investment and a focus on profitability.
Mundo AI secures $20M in Series A funding to build datasets, evaluations, and research for AI perception across audio, video, and emerging modalities.