JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Training and evaluation guide

kev

Kev documents a local typed-decision server, its Qwen training recipe, and probes for batching, option-order sensitivity, question isolation, and delimiter handling.

Original by Jared PalmerEvaluationGitHub repositorySource reviewed

Before you dive in

What you’ll find in the original

  1. Reproduce the released model family's base recipe from 10,000 public examples, 896 generated policy examples, and 1,680 generated rule examples.
  2. Compare parallel and sequential questions, permute every six-option ordering, and probe whether sibling questions leak information.
  3. Escape delimiter-like user strings and distinguish serialized answer-token accounting from generated tokens.
Worth knowing

The README reports the recipe and probe behavior for local Kev models; JevMade did not train or run them.

More guides →See the experiments →