@xiciyang/training-sft

Same capability topic

Related by capability

Ranked by shared capability topic, runtime face, tools or UI contributions, and distribution mode. This is not a similarity, quality, or compatibility ranking.

• @xiciyang/training-pretrain继续预训练（continued pretraining）薄插件：领域语料校验（双轨去泄漏 + 全量 PII）→ 公共训练循环（training-core，容器内 transformers）→ 分片续训（步号与数据位置连续性机器可验）→ 合并 → serving 部署 → 遗忘门禁。依赖长任务 1 冻结的六个 @ai-infra/* 共享库。Bundles & other modules · Installable
• @xiciyang/training-rlRL（GRPO）薄插件：规则奖励（显式版本）→ β 可推导（β=σ_r/KL*，σ_r 实测）→ training-core 单体档位驱动 TRL GRPOTrainer → 产物级断言 → 合并 → serving 部署 → 偏好对齐 + 遗忘门禁。依赖长任务 1 冻结的六个 @ai-infra/* 共享库。Bundles & other modules · Installable
• dsh-session-foldersGive every new DSH conversation its own dated directory: new sessions created in a configured root workspace are redirected into <root>/<YYYY-MM-DD-HH-mm-ss>/, which is registered as its own workspace so files never mix.Bundles & other modules · Installable
• toolshrinkReduce large agent tool output by what it means, not by where it was cut.Bundles & other modules · Installable
• dsh-session-isolationWorkspace session isolation for DeepSeek Harness: give every session its own output directory and refuse to overwrite files owned by another live session in the same workspace.Bundles & other modules · Installable
• @towzai/dsh-memoryDSH 跨会话记忆插件 | Cross-session memory plugin: YAML 存储、embedding 语义检索、标题优先注入、两级自动注入Bundles & other modules · Installable