Skip to content

Contributing — Mens native training

This guide covers how to contribute to the vox mens train pipeline and how to ensure your code changes feed the MENS training loop correctly.

SurfaceLocation
CLIvox mens traincrates/vox-cli/src/commands/schola/train/
Libraryvox_populi::mens::tensor::run_mens_training (lora_train.rs)
ContractFineTuneContract, ExecutionPlanner, preflight_train

MENS training relies on high-quality .vox examples. When you add or modify code in examples/golden/*.vox, it must pass validation before being ingested.

The flow is:

  1. examples/golden/*.vox (human or agent written)
  2. vox corpus validate-batch (CI pipeline check)
  3. Output: golden_validated.jsonl (ready for training)

To ensure your code becomes a positive training example:

  • Parse Rate: The code must pass the parser 100% cleanly. vox corpus eval --mode ast must succeed.
  • Test Blocks: Use @test blocks to validate logic. The AST coverage and test pass rates will be used in future GRPO reward shaping.
  • No Stubs: Zero todo!(), unimplemented!(), or empty function bodies.
  • Rich Constructs: Use idiomatically correct Vox patterns.

Run the training planner tests:

Terminal window
cargo check -p vox-populi --features mens-train
cargo test -p vox-populi --features mens-train execution_planner

Validate the golden corpus locally:

Terminal window
cargo run -p vox-cli -- corpus eval --mode ast examples/golden/

A PR contributing to the MENS pipeline is “done” when:

  • cargo test -p vox-populi --features mens-train is green.
  • No new parse failures are introduced to the golden corpus.
  • TOESTUB reports zero stubs or god-object violations in the touched code.
  • Any new CLI flags are documented in the SSOT.

Candle QLoRA is the active vox mens train backend; keep docs and error messages aligned (lora_train.rs is authoritative when in doubt).