JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Training and gameplay benchmark

NanoJev

NanoJev describes a local 0.6B parallel decision model, fixed data splits, replay-checked held-out runs, and comparisons across Maze, Snake, ViZDoom Basic, and ViZDoom Predict Position.

Original by TianyuCodingsEvaluationGitHub repositorySource reviewed

Before you dive in

What you’ll find in the original

  1. Keep train, development, calibration, test, and out-of-distribution partitions explicit across 18,760 questions per target variant.
  2. Reuse the same checkpoint across tasks so results test transfer rather than per-game retraining.
  3. Publish replay-checked trajectories and attempt or horizon counts, not only selected screenshots.
Worth knowing

The author reports 16,333 ViZDoom questions, a 225-attempt showcase maze run, and 30 food items in a 256-step Snake showcase; the separate held-out table reports outcomes over 274 test cases. JevMade did not rerun either set.