Dictation Format Adapters

This project implements a post-processing layer for speech-to-text models, utilizing a base 0.6B parameter model to refine raw dictation. It employs a bounded catalogue approach, where specific LoRA adapters are applied to the base model to output text in predefined shapes rather than relying on open-ended instruction following. This method ensures high reliability for routine tasks like formatting emails, blog outlines, or task lists.

The development pipeline involves a data-distillation process using teacher prompts from a larger model, followed by deterministic shape checks to validate output quality. By stacking specialized LoRAs onto a general-purpose cleanup base, the system provides an efficient, modular solution for users who require consistent, structured text from voice inputs without the overhead of larger language models.