machine
Parametric learning in your code harness.
Star
Fork
Watch

Simplify Machine to one trace pool and honest training #18

Open cameron.stream
Open opened by and assigned to cameron.stream
targeting main

Machine collected four JSONL files, two training CLIs, and a Tinker wizard. This branch makes the product loop one thing: append a trace, train the good ones, keep one sampler.

  • /good / /bad / /machine auto all write traces.jsonl. Every row is a full last-turn trajectory. /train uses tag: "good" only.
  • Scaffolds and eval artifacts go in the project store, not the working tree.
  • /tinker only loads checkpoints (TUI picker, use, import from a log dir). Wizard/demo/deploy and the duplicated SFT loop are gone.
  • /train run is a real Tinker job (--steps 2 for a short run). /train smoke is removed.
  • After a run, Machine keeps the new sampler and one rollback, and deletes training-state weights (Cookbook would keep finals forever). /tinker prune lists account leftovers.

Euler is unchanged.

Commits 3
All changes
Remove /train smoke; keep a rollback sampler after each run.
Keep one hosted sampler per project; rename /train smoke to /train run.
Simplify to a single trace pool and harden training and validation