Unverified50% confidenceBenchmarkExact time
在175张训练图/132张测试图的细粒度大众高尔夫车型分类任务中，SigLIP2 SO400M的k-NN准确率约为92%
1
Sources
50%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
DINOv2 vs SigLIP：k-NN分类为何差距悬殊？视觉编码器选型避坑指南
redditr/MachineLearning7/8/2026
Related Claims
Unverified在细粒度汽车分类的k-NN实验中，SigLIP2 SO400M达到约92%准确率，CLIP ViT-L约59%，DINOv2 Giant约41%73% similarUnverifiedBullet 在 SWE-bench Verified 基准测试中取得 95.8% 的成绩，位列前三，平均每个任务耗时 119 秒65% similarUnverified在同一细粒度汽车分类任务中，DINOv2 Giant的k-NN准确率约为41%，表现垫底59% similarUnverified在同一细粒度汽车分类任务中，CLIP ViT-L的k-NN准确率约为59%56% similarUnverifiedPathway报告称基于BDH的系统在约25万道Sudoku Extreme题目上达到97.4%的Top-1准确率，且未使用CoT、未进行解答回溯56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/456341API
curl https://kongchang.com/api/v1/knowledge/claims/456341MCP
get_claim(id=456341)