Training
Train models through the Tinker-compatible SDK endpoint, then inspect jobs and checkpoints with the released el CLI. EnvPlatform manages the training backend and resources.
Choose a workflow
| Goal | Start here |
|---|---|
| Run supervised fine-tuning (SFT) | SFT quickstart |
| Save weights or resume training | Checkpoints and recovery |
| Query jobs and estimated usage | Jobs and billing |
Training concepts
- Train Job records training status and ownership under
train_job_id. It is separate from an Agent Job, Task or Trial. - Session groups SDK operations and backend resources under
session_id. Creating a gateway session also creates its Train Job record. - Training client holds trainable model state. The script submits gradient calculations, optimizer steps and checkpoint operations through this client.
- Policy is an administrator-approved configuration bound to the workspace and user. It fixes permitted models, resources, budget and deadline. Changing a script's model argument does not create a policy or override it.
- Checkpoint preserves training state or sampler weights. Its owner and purpose determine how it can be used.
The SDK script controls the training loop. Keep it running until its operations and finish request complete; closing the terminal does not establish a successful training outcome. Agent rollout or evaluation uses separate Agent resources, with separate IDs.
Available interfaces
| Interface | Current behavior |
|---|---|
Released envloop-cli 0.1.1 | el train job list/status, checkpoint list, billing usage |
| envloop source 0.1.0 | envloop.training re-exports Tinker 0.30.2; source CLI provides el train submit/status |
| Tinker-compatible SDK | Creates sessions and models; runs training, checkpoint and optional sampling operations at /tinker |
| Managed Train Job API | /v1/train/jobs submission, result and cancellation require a configured managed TrainJobService; read-only gateway records do not prove those write operations are enabled |
envloop.Client has no client.train property. Import training separately with from envloop import training; this preserves Tinker's methods, futures and exceptions. training.TrainingRun is a data type, not a lifecycle manager. Do not assume the published CLI includes the source SDK or the source submission commands.
Verification scope
Checked on 2026-10-11 against EnvPlatform 7fcf9ed8 and envloop 39ecb502. The recorded GPU smoke completed actor-only LoRA training and checkpoint verification on 2×H200. It does not establish full-dataset quality, billing accuracy or sampling on two GPUs. This documentation change does not rerun GPU training.