Cookbook

The cookbook contains complete examples for common training tasks. Use them to run supervised or reinforcement training against the trainer at TRAINER_URL. Each example directory contains data preparation, descriptors, and walkthrough.sh, which provides the akka commands. You need Python 3, the CLI, and a trainer. Each page lists any other requirements and links to an archive of its example directory.

Choose an example based on the method and features that you want to try:

Example	Method	Shows
Support triage	Supervised	Narrow classification graded by exact match, and the tuned model compared with an untuned baseline on the same rows.
GSM8K	Reinforcement	A reward function in a scoring bundle, a format term combined with a correctness term, and the same function used as the grader.
Spider text to SQL	Supervised, then reinforcement	A five-stage pipeline, a scoring bundle that includes the databases it queries, and reinforcement learning that improves on supervised tuning for unseen databases.

Example

Method

Shows

Support triage

Supervised

Narrow classification graded by exact match, and the tuned model compared with an untuned baseline on the same rows.

GSM8K

Reinforcement

A reward function in a scoring bundle, a format term combined with a correctness term, and the same function used as the grader.

Spider text to SQL

Supervised, then reinforcement

A five-stage pipeline, a scoring bundle that includes the databases it queries, and reinforcement learning that improves on supervised tuning for unseen databases.

Every walkthrough reads ROWS and EVAL_ROWS from the environment. The Spider walkthrough also reads RL_ROWS. Use the suggested values on each page for a small first run and an estimate of its duration.