Workspace connections and training reliability
Workspace SSH connections now use the node identity expected by the gateway. CPU workspaces that fit an organization’s shared node can use it; larger disk requests retain dedicated placement.
Deleting a stopped sandbox no longer stalls when another status update is finishing.
TrainingJob sweeps validate each cell, enforce the parent’s spending cap, and retain final cost information. Missing model revisions report the revision error. Local estimates include a declared run duration.
The managed SFT, DPO, and distillation runtimes include CPU presets for the tiny Qwen3 test model in Full and LoRA modes. Other models keep their existing presets, and environment training still requires a GPU host.
Managed training runtimes and dataset previews accept dataset paths from the Volume root, such as
/train.jsonl, as well as relative paths. Paths that escape the Volume remain rejected.