Commit Graph

114 Commits

Author SHA1 Message Date
ViperEkura e52803ddc3 refactor(trainer): 将回调类移至独立文件并优化训练器结构 2025-09-29 12:00:25 +08:00
ViperEkura 8206c7855e fix(transformer): 调整注意力掩码处理逻辑 2025-09-29 11:31:42 +08:00
ViperEkura 816bc78894 feat(trainer): 引入训练器回调机制并重构训练流程 2025-09-29 11:31:31 +08:00
ViperEkura 92999fa9f6 fix(trainer): 修复训练器中配置引用错误的问题 2025-09-28 22:20:25 +08:00
ViperEkura 1c9063fd3d refactor(trainer): 统一参数命名以提升可读性 2025-09-28 22:14:24 +08:00
ViperEkura fa43ed2943 feat(trainer): 重构训练配置与策略工厂引入 2025-09-28 21:39:48 +08:00
ViperEkura 1169cfad82 fix(trainer): 修复多轮对话中的因果注意力掩码计算逻辑等 2025-09-28 15:15:19 +08:00
ViperEkura 25ec56a1f5 fix(trainer): 修复训练器恢复检查点时的学习率初始化问题 2025-09-28 14:38:02 +08:00
ViperEkura f25a249291 feat(khaosz): 优化模型参数保存与加载逻辑 2025-09-28 14:00:21 +08:00
ViperEkura 4fcdc87c95 feat(trainer): 重构数据集与策略模块以支持字典形式的数据返回 2025-09-27 14:11:27 +08:00
ViperEkura 9fbc9481b5 refactor(core): 修改注意力掩码处理函数并重命名参数 2025-09-27 13:37:10 +08:00
ViperEkura 053f4a4dad feat( StrategyFactory): 添加 SFT 策略初始化参数并完善工厂方法调用 2025-09-27 13:24:16 +08:00
ViperEkura 676fdd59d7 feat(strategy): 重构mask构建逻辑并优化策略工厂参数传递 2025-09-27 13:12:57 +08:00
ViperEkura a4443765ee Initial commit 2025-09-27 12:02:22 +08:00