home.social

#ocrbench — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #ocrbench, aggregated by home.social.

fetched live
  1. Alibaba’s new open‑source model Qwen3‑VL can scan two‑hour videos, achieving 96.5 % on DocVQA and 875 on OCRBench. The multimodal vision‑language system rivals the rumored GPT‑5 in document understanding. Dive into the results and see why the community is buzzing. #Qwen3VL #Alibaba #DocVQA #OCRBench

    🔗 aidailypost.com/news/qwen3vl-s

  2. #NVIDIA introduces #NVLM 1.0, a family of open-source #multimodal #LLMs:

    🏆 Achieves state-of-the-art results on vision-language tasks, competing with #GPT4 and #Llama3V

    📊 72B model outperforms on #OCRBench and #VQAv2 benchmarks

    📈 Shows improved accuracy on text-only tasks after multimodal training

    💻 Excels in #math, #coding, and #reasoning across modalities

    🧠 Novel architecture enhances training efficiency and multimodal reasoning

    🖼️ Introduces 1-D tile-tagging for improved performance on high-resolution images

    🔬 Emphasizes dataset quality and task diversity over scale in training

    🔗 Open-sourcing model weights and training code in Megatron-Core

    Learn more: research.nvidia.com/labs/adlr/

  3. #NVIDIA introduces #NVLM 1.0, a family of open-source #multimodal #LLMs:

    🏆 Achieves state-of-the-art results on vision-language tasks, competing with #GPT4 and #Llama3V

    📊 72B model outperforms on #OCRBench and #VQAv2 benchmarks

    📈 Shows improved accuracy on text-only tasks after multimodal training

    💻 Excels in #math, #coding, and #reasoning across modalities

    🧠 Novel architecture enhances training efficiency and multimodal reasoning

    🖼️ Introduces 1-D tile-tagging for improved performance on high-resolution images

    🔬 Emphasizes dataset quality and task diversity over scale in training

    🔗 Open-sourcing model weights and training code in Megatron-Core

    Learn more: research.nvidia.com/labs/adlr/