Skip to main content

​
Training and selection limits

A project dataset can contain more rows than a Training selection supports. With the default split, a dataset of more than 125,000 rows puts over 100,000 rows in its train part, and the estimate is blocked. To train on fewer rows, freeze a smaller slice with toloka dataset slice create <dataset> --filter <field>=<value> and pass it with --slice, or split the dataset into several train parts with toloka dataset split and pass only some of them with --parts. See Slices and split columns. Training reads train and validation parts from one frozen slice. If the slice supplies no validation rows, the platform holds out 10% of the selected train rows, rounded up, and checks that the remaining rows can fill one batch. A fresh 10-row dataset with the default 80/10/10 split provides 8 training rows, 1 validation row, and 1 test row. See Dataset selections. The estimate blocks an example above maxSequenceLength; it does not truncate it.

​
Gisting prompt limits

The CLI validates the local prompt before uploading it. The Training Run snapshots the prompt and refuses to publish if it changed after upload.

​
Cost estimate limits

An estimate is open until it is consumed by a Training Run or expires.

​
Training time limit

A ready estimate’s maximum training wall time must be six hours or less. This limit applies to the estimated training window, not the total time until the Training Run completes. The surrounding lifecycle can also include data preparation, a GPU-capacity wait, output validation, billing, and publication. A job that trains longer than estimated remains capped at the approved maximum charge and can still publish its Model Version. For one-GPU SFT models, a duration issue can sometimes be resolved by choosing the eight-GPU qwen3-32b or qwen3-235b. Otherwise reduce rows or epochs. Gisting supports only the one-GPU tiers, so reduce rows or epochs.

​
Model Version limits

The CLI requests signed links and verifies every downloaded file’s size and SHA-256. If a link expires, rerun the download command to request fresh links. Gisting is limited to qwen3-4b and qwen3-8b because the larger models’ full weights cannot fit the Model Version limits. SFT publishes adapters and supports all catalog tiers.

​
Payment window

A Training Run blocked while collecting its final charge has seven days to complete payment. Add funds and resume within that window. After the deadline it fails with payment_window_expired, cannot resume, and is not charged.