"Deprecation policy — Mens native fine-tuning"

Deprecation policy — Mens native fine-tuning

Stable

  • vox mens train with --backend lora and --backend qlora.
  • vox schola merge-qlora (alias merge-adapter).
  • vox mens merge-weights for Burn *.bin LoRA checkpoints.

Deprecated / transitional

  • vox train --native-lora: use vox mens train --backend lora (stderr deprecation already emitted from dispatch).
  • Backend-only mental model: prefer the contract fields (tokenizer mode, quant mode, adapter method) when scripting; CLI flags remain the user-facing surface until a preset/JSON contract ships.

Timeline

  • No CLI flags removed in this iteration; aliases added (merge-adapter).
  • Future removal of legacy paths will be announced in this doc + mens-training.md with one release notice.