inclusionAI Releases ZoomGain Models, Benchmark and RL Data

inclusionAI released the ZoomGain arXiv paper, the ZoomGainBench evaluation benchmark, an RL training dataset, and ZoomGain-4B and ZoomGain-9B vision-language model checkpoints. The method mines global-to-zoom perception gaps to build fine-grained visual supervision, and inclusionAI says ZoomGain-9B reached superior performance on multiple fine-grained and general perception benchmarks. ZoomGainBench contains 1,000 questions over 941 images, and the RL dataset holds 9,000 training records drawn from 7,713 source images. The best of 28 models scored 61.08 on the benchmark, according to inclusionAI. The code is released under the Apache License 2.0, while the bundled training framework keeps its upstream license and notices. Users may need to run hf auth login if the Hugging Face repositories require authenticated access, and the terms of each dataset and checkpoint are set out on their Hugging Face cards.

สร้างจากรายงาน · อัปเดต 14 ชม. ที่แล้ว

ยังมีการเฝ้าดูต่อเนื่องไม่พอจะวาดแนวโน้ม