songsenand
0 Followers · 0 Following
  • Joined on 2024-01-17
Repositories
11
 Projects Packages Public Activity Starred Repositories
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-07-07 16:24:02 +08:00
d2c94749cb fix(input processing): 增加序列长度一致性校验避免模型溢出
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-25 14:04:16 +08:00
1e77c55ece feat(prepare_inputs): 添加 part4_ids 构建逻辑并作为输入参数传递
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-25 08:22:26 +08:00
079cbd21a1 feat: 添加 merge-shuffle 模式支持低磁盘占用打乱
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-24 09:44:39 +08:00
be8eaa768c refactor(model): 从checkpoint中动态加载模型配置和权重
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-23 18:29:02 +08:00
3cd912c0d0 feat(trainer): 重定向日志输出以避免干扰进度条显示
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-23 14:24:16 +08:00
2159d8df1b fix(trainer): 优化 part4_ids 处理逻辑，避免重复获取和设备转移
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-23 13:58:54 +08:00
273afe7fe1 feat(data): 添加标签补充与平衡子采样工具及文档
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-19 21:27:33 +08:00
5bc8ac0480 feat(docs): 添加 Part4 候选词提示的详细设计与训练流程说明
songsenand created branch feat/confusion-finetune in songsenand/SUimeModelTraner 2026-06-17 09:02:26 +08:00
songsenand pushed to feat/confusion-finetune at songsenand/SUimeModelTraner 2026-06-17 09:02:26 +08:00
bfff409b35 feat(confusion-finetune): 引入拼音混淆样本生成与三路交叉注意力架构
songsenand deleted branch numresblock18 from songsenand/SUimeModelTraner 2026-06-01 14:51:26 +08:00
songsenand pushed to main at songsenand/SUimeModelTraner 2026-06-01 14:50:01 +08:00
88955bcfdd refactor(model): 优化slot权重计算逻辑以提升稳定性
53f244de2f docs: 详细描述history_slot_ids的设计策略与使用场景
71ef54e3d4 fix(trainer): 使用固定最大序列长度的collate函数以避免内存问题
722912f296 feat(data-preprocess): 预处理数据预打乱以提升训练效率
0862b5b8fc fix(PreProcessedDataset): 修复数据类型转换，避免内存复制
Compare 15 commits »
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-23 13:42:54 +08:00
88955bcfdd refactor(model): 优化slot权重计算逻辑以提升稳定性
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-23 13:25:43 +08:00
53f244de2f docs: 详细描述history_slot_ids的设计策略与使用场景
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-15 14:47:42 +08:00
71ef54e3d4 fix(trainer): 使用固定最大序列长度的collate函数以避免内存问题
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-15 13:49:49 +08:00
722912f296 feat(data-preprocess): 预处理数据预打乱以提升训练效率
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-11 22:05:16 +08:00
0862b5b8fc fix(PreProcessedDataset): 修复数据类型转换，避免内存复制
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-11 00:14:31 +08:00
27beb7f0b1 refactor(trainer): 优化进度条逻辑与训练循环结构
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-10 23:34:17 +08:00
d0f1534086 fix(dataset): 修复分片数据集时未正确计算样本数的问题
songsenand pushed to numresblock18 at songsenand/SUimeModelTraner 2026-05-10 10:38:25 +08:00
483e4d4f98 fix(model): 移除 torch.compile 的注释和未使用配置