Training APIValidate, create, queue, monitor, stop, and resume fine-tuning jobs. All wire fields use canonical snake_case. Read capabilities and run preflight before creating a paid job.
Managed training. You choose the model, adapter, data and training settings (learning rate, epochs, context length, warmup, weight decay, schedule, precision, optimizer, LoRA rank and alpha, validation rows, early stopping). We choose the machines, batch size and memory settings from setups tested for that model, so a run is as fast as it can be and does not run out of memory. GPU, batch and checkpoint-count fields are ignored with a FIELD_IGNORED warning. You pay per trained token: preflight returns the estimate and the job is charged for the exact tokens it trained. When no machine is free the job waits in the queue, first come first served and unbilled, starts automatically, emails you when it starts, and can be cancelled at any time. Customers see training metrics (loss, eval loss, gradient norm, accuracy, learning rate, epoch, tokens trained); logs and machine metrics are not part of the API. GET /api/training/managed-options lists the precision, optimizer and context lengths offered for a model and adapter.
GET /api/training/capabilitiesGet the authoritative contract for the currently compatible pinned BIOS image Get the authoritative contract for the currently compatible pinned BIOS image GET /api/training/gpu-optionsLegacy GPU-time training only. Under managed training this endpoint answers 404: the platform chooses the machines Legacy GPU-time training only. Under managed training this endpoint answers 404: the platform chooses the machines POST /api/training/preflightValidate and canonicalize the exact request without charging, queueing, creating a job, or allocating a GPU Validate and canonicalize the exact request without charging, queueing, creating a job, or allocating a GPU POST /api/training/jobsCreate the exact paid request reviewed through preflight. Safe retries must reuse the same idempotency key and payload. Create the exact paid request reviewed through preflight. Safe retries must reuse the same idempotency key and payload. GET /api/training/jobsList training jobs with server-side filtering and pagination List training jobs with server-side filtering and pagination GET /api/training/jobs/:idGet detailed lifecycle, stage, progress, billing, stop, recovery, and artifact state Get detailed lifecycle, stage, progress, billing, stop, recovery, and artifact state POST /api/training/jobs/:id/stopPersist a durable stop intent, preserve data when requested, stop billing at request time, and reconcile until the GPU is released Persist a durable stop intent, preserve data when requested, stop billing at request time, and reconcile until the GPU is released POST /api/training/jobs/:id/resumeCreate a new child job from the frozen last committed checkpoint and original training request. A job trained on top of earlier versions (start_from) cannot be resumed: 422 START_FROM_RESUME_UNSUPPORTED, start it again instead. Create a new child job from the frozen last committed checkpoint and original training request. A job trained on top of earlier versions (start_from) cannot be resumed: 422 START_FROM_RESUME_UNSUPPORTED, start it again instead. GET /api/training/jobs/:id/metricsGet training metrics (loss, learning rate, etc.) Get training metrics (loss, learning rate, etc.) GET /api/training/jobs/:id/logsGet training logs (stdout/stderr) Get training logs (stdout/stderr) GET /api/training/jobs/:id/checkpointsList saved checkpoints for a training job. format says what each is: adapter, full, or merged -- the base with a start_from job's adapters and its selected adapter merged into one model, named merged model (step B), with step the adapter step it was merged from and merged_of_checkpoint_id that adapter; a merged model is never final, best or pruned. An adapter of a job with start_from is not deployable on its own (deployments answer 400 CHAINED_ADAPTER_NOT_DEPLOYABLE): deploy its merged model instead. List saved checkpoints for a training job. format says what each is: adapter, full, or merged -- the base with a start_from job's adapters and its selected adapter merged into one model, named merged model (step B), with step the adapter step it was merged from and merged_of_checkpoint_id that adapter; a merged model is never final, best or pruned. An adapter of a job with start_from is not deployable on its own (deployments answer 400 CHAINED_ADAPTER_NOT_DEPLOYABLE): deploy its merged model instead. DELETE /api/training/jobs/:jobId/checkpoints/:cpIdDelete a specific checkpoint. 409 CHECKPOINT_IN_USE while a live pipeline version, or one a rollback can still put back, answers from it, or a running job trains on top of it: delete the pipeline or roll the version back first. Delete a specific checkpoint. 409 CHECKPOINT_IN_USE while a live pipeline version, or one a rollback can still put back, answers from it, or a running job trains on top of it: delete the pipeline or roll the version back first. Run BiOS Documentation. Need help? Email contact@runbios.ai