开源决策模型 basal-omni-ultra 在 Werdykt v1 上跑赢 Jev 1.14
作者自建 5000 题隐藏测试，basal-omni-ultra 宏准确率 0.879 对 Jev 1.14 的 0.822，10 个领域赢 9 个，榜单公开待独立复测。
Remek Kinas (@KinasRemek), author of the open basal decision models, showed that his basal-omni-ultra scores 0.879 macro accuracy on the Werdykt v1 test against 0.822 for TypeSafe’s Jev 1.14, and leads in 9 of 10 areas. The model handles text, image and sound in Polish and English, plus an extended query schema Jev does not have (multi, act, tool), and it scores 0.928 on agent and tool decisions. Someone will ask whether this is the author’s own test, and yes, but it consists of 5,000 hidden items, the same ones go to every model, each gets one call per item, and the leaderboard and samples are public at http://basal.si5.pl. The release is on Saturday, so I am waiting for independent measurements. In the one area where it does not lead, real documents, everyone sits around 0.99, so it is a tie up to the rubber stamp.
这条是引用着另一条写的：x.com/kinasremek/status/2110651784033456229