Training jobs

Fine-tune an open-weight model from the dashboard: creating jobs, monitoring runs, and using the results.

A training job fine-tunes an open-weight base model with LoRA on Arkor's managed GPUs. Jobs live under a project's Jobs tab. If you prefer code over forms, the framework docs cover the same platform through the TypeScript SDK and CLI.

Creating a job

Jobs → New Job takes:

Field	Notes
Job name	Free text, for the job list
Model	Hugging Face repo id of the base model
Dataset	Hugging Face dataset id
Dataset format	How the dataset's columns map to training text: Alpaca (`instruction` / `input` / `output`), ChatML (a `messages` list), ShareGPT, prompt-completion, plain text, or pre-tokenized
Max samples	Optional cap on dataset rows; empty trains on all rows
Max steps	Optional cap on optimizer steps; empty trains by epochs
Learning rate	Defaults to `0.0002`

Submitting queues the job. Arkor picks a GPU from its provider pool and dispatches the run; you do not choose or manage GPUs.

Monitoring a run

The job detail page shows:

Status — queued, running, preempted, completed, failed, or cancelled, with a Cancel button while queued, running, or preempted. preempted means the GPU the job was training on was reclaimed mid-run; the job is not lost — it waits to be picked up again and resumes from its last checkpoint automatically.
Latest metrics — the step, loss, and epoch from the newest trainer event at load time; reload the page for fresh numbers.
Training events — the event log as a table: per-step log lines with step, loss, and time, plus lifecycle events.
Configuration — the job id, provider, timestamps, and the full submitted config as JSON.

Failures surface the trainer's error message in a banner.

A completed job's detail page with metrics and the event log

Checkpoints and adapters

During the run the trainer uploads mid-run checkpoints; on success it uploads the final adapter. Both become usable in two places:

The project Playground can chat with the final adapter or any checkpoint of a completed job.
An endpoint can serve the final adapter or a specific checkpoint at a stable *.arkor.app URL. The endpoint's target can be swapped in place through the SDK's deployments API, so you can promote a better checkpoint without changing the URL your app calls.

A completed job's detail page links straight to the Playground preloaded with that job.

License

The content of this page is licensed under CC BY 4.0.

Code samples are licensed under the MIT License, which asks you to keep its copyright and permission notice in all copies.

Training jobs

Fine-tune an open-weight model from the dashboard: creating jobs, monitoring runs, and using the results.

Creating a job

Jobs → New Job takes:

Field	Notes
Job name	Free text, for the job list
Model	Hugging Face repo id of the base model
Dataset	Hugging Face dataset id
Dataset format	How the dataset's columns map to training text: Alpaca (`instruction` / `input` / `output`), ChatML (a `messages` list), ShareGPT, prompt-completion, plain text, or pre-tokenized
Max samples	Optional cap on dataset rows; empty trains on all rows
Max steps	Optional cap on optimizer steps; empty trains by epochs
Learning rate	Defaults to `0.0002`

Submitting queues the job. Arkor picks a GPU from its provider pool and dispatches the run; you do not choose or manage GPUs.

Monitoring a run

The job detail page shows:

Status — queued, running, preempted, completed, failed, or cancelled, with a Cancel button while queued, running, or preempted. preempted means the GPU the job was training on was reclaimed mid-run; the job is not lost — it waits to be picked up again and resumes from its last checkpoint automatically.

Latest metrics — the step, loss, and epoch from the newest trainer event at load time; reload the page for fresh numbers.

Training events — the event log as a table: per-step log lines with step, loss, and time, plus lifecycle events.

Configuration — the job id, provider, timestamps, and the full submitted config as JSON.

Failures surface the trainer's error message in a banner.

Checkpoints and adapters

During the run the trainer uploads mid-run checkpoints; on success it uploads the final adapter. Both become usable in two places:

The project Playground can chat with the final adapter or any checkpoint of a completed job.

An endpoint can serve the final adapter or a specific checkpoint at a stable *.arkor.app URL. The endpoint's target can be swapped in place through the SDK's deployments API, so you can promote a better checkpoint without changing the URL your app calls.

A completed job's detail page links straight to the Playground preloaded with that job.

License

The content of this page is licensed under CC BY 4.0.

Code samples are licensed under the MIT License, which asks you to keep its copyright and permission notice in all copies.

Training jobs

Creating a job#

Monitoring a run#

Checkpoints and adapters#

Training jobs

Creating a job#

Monitoring a run#

Checkpoints and adapters#

Creating a job

Monitoring a run

Checkpoints and adapters

Creating a job

Monitoring a run

Checkpoints and adapters