Uratori updates a compact model for Japanese evidence checks
Uratori returns evidence judgments without generating an answer.
Creator tokimoa released retrained Uratori 310M weights on October 6, following an October 5 debut. The model card describes a Japanese model that scores supplied evidence rather than generating text.
It targets grounding checks, document comparisons and retrieval judgments. Downloadable weights use Apache 2.0 licensing. A 316 MB ONNX version supports CPU use.
Accuracy needs context
The creator reports 68.6% accuracy on 802 test items. Labels combine model judges with intended answers, rather than human annotations. That leaves roughly one wrong judgment in three against those labels.
Inputs are limited to 1,024 tokens. The model checks supplied Japanese text, not facts against outside knowledge. Its documentation excludes sole reliance for consequential decisions. ByteForward has not run the model.
Related reporting examines why uncertainty warnings need reliable targeting.
Illustrative wooden puzzle photograph by Gorkaazk under CC0. Converted to WebP. These unrelated objects illustrate comparison and are not model outputs.
