JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Pipeline guide

Screen data, train a model, and evaluate the change

A traceable data workflow from upload and local checks through Jev quality screening, model training, and before-and-after evaluation, with browser, CLI, Python, and HTTP entry points.

Original by RenaGaoEvaluationGitHub README and screening moduleSource reviewed

Before you dive in

What you’ll find in the original

  1. Run format, length, and exact-duplicate checks locally before asking Jev to partition rows into keep, review, and reject sets.
  2. Treat Jev as the data evaluator: the configured Hugging Face model, not Jev, is what subsequent fine-tuning updates.
  3. Export portable training bundles and compare measured results after SFT, GRPO, or PPO rather than treating successful training as quality evidence.
Worth knowing

The public demo uses local rules and a byte-bigram model, not Jev or LLM training, and is limited to 2 MiB or 1,000 rows. Full training needs self-hosted compute; GPU requirements vary by path.

More guides →See the experiments →