How-to guides¶

These are task-oriented recipes for getting specific things done. They assume you have the package installed or, for the GPU jobs, a Modal account. For end-to-end walkthroughs, see the Tutorials instead.

Guides¶

• Fine-tune on your own dataset: write your labelled questions as JSONL and train the checkpoint on them, locally or on Modal.
• Calibrate on your own data: fit per-type temperatures on your labelled questions and check that they help.
• Run jobs on Modal: the volumes and secret modal_app.py expects, and the commands for data preparation, training, evaluation and publishing.
• Evaluate a checkpoint: run the whole evaluation suite on one checkpoint and turn the results into a scorecard.
• Play games with a checkpoint: watch a checkpoint play in a local window, and score it on the games suite.

See also¶

• Concepts: the reasoning behind these tasks.
• Reference: every argument, field and file format.