Chinese Brief
中文案例导读
Sioonn 公开的代码代理与软件工程案例,来源为 公开代码库,复核于 2026-06-27T08:18:08Z。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
CASE EVIDENCE / A RECORD
Sioonn 公开的代码代理与软件工程案例,来源为 公开代码库,复核于 2026-06-27T08:18:08Z。
原始记录:K-fashion text-to-image retrieval project uses Solar Pro3 to generate Korean user-style synthetic queries
Chinese Brief
Sioonn 公开的代码代理与软件工程案例,来源为 公开代码库,复核于 2026-06-27T08:18:08Z。 Model Atlas 将它标记为 A 类证据,因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
任务
这是一个围绕软件工程任务执行的真实任务,公开材料可以回溯到具体使用者和具体产物。 原始资料写作:Improve Korean natural-language fashion image search by generating user-style Korean search queries from product captions and using them to align a SigLIP 2 text tower with fashion product images.
公开材料提供公开代码、README 或项目配置,可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作:The project reports a lightweight retrieval pipeline where MRR improves from 0.571 zero-shot SigLIP 2 to 0.760 after text-only LoRA, BM25 fusion and query-bank prior correction on human gold queries.
Solar Pro 3 在该案例中承担软件工程任务执行相关的生成、分析、编排或实现角色。 原始资料写作:Solar-pro3 is used as the external API model for training-query generation, creating three Korean user-style synthetic queries per product caption for a 7,023-query bank used in LoRA alignment and prior correction.
当前判断基于公开材料;若产物下线、仓库变更或模型参与比例仅来自作者自述,需要在引用前重新复核。 原始资料写作:Evidence is self-declared in a public research-code README; the reported metrics are project results and should be treated as repository-authored claims.