The cleanup model: a fine-tune we trained for dictation

After transcription, Evoglyph turns raw speech into clean, readable text with a language model it trained for exactly this job. It is a small fine-tune layered on an openly licensed base model, it runs entirely on your Mac, and it comes with a hard rule: it can tidy your words, but it can never make you say something you didn't. This article covers what runs, why it is a fine-tune rather than a prompt, and where it came from.

What runs on your Mac

Cleanup is two files working together. The base is Liquid AI's LFM2-2.6B, a general language model, shipped as a 4-bit build that is about 1.45 GB on disk. On top of it sits Evoglyph's own adapter, a small low-rank file of about 23 MB that changes how the base behaves without changing the base itself. The base stays fixed; the adapter is the part Evoglyph trained.

Both run in-process on the Apple GPU through MLX. There is no separate service, no network call at inference time, and nothing about your text leaves the device during cleanup.

Why a fine-tune, not just a prompt

A general-purpose model is a strong generalist and a mediocre dictation cleaner. Ask it to "tidy this up" with a prompt and the results wander: it sometimes rephrases, sometimes summarises, sometimes changes a word you meant. For text that lands directly in your commit messages and documents, that inconsistency is the problem, not a detail.

Training a small adapter for this one job makes the behaviour consistent and narrow: recover fillers, merge stutters, fix homophones from context, and stop there. It also means Evoglyph can ship the exact prompt the adapter was trained under, so the model you run is the model that was tuned, not a general model coaxed with instructions.

What it does to your words

The cleanup model takes flat, raw transcription and:

removes filler, false starts, and stutters;
collapses repetition to a single instance, including emphatic repetition, so "no no no" becomes "No";
normalises casual speech into readable text while keeping your abbreviations intact;
joins spelled-out digits and acronyms;
fixes grammar; and
corrects context homophones (their versus there, one versus won). That is a skill trained into this same adapter; see homophone correction for how it works.

Through all of that, it preserves your meaning. It does not paraphrase and it does not summarise. The words that come out are your words, tidied.

What it will never do

Cleanup can never block your text. If it fails, if it runs past its time budget, or if the safety net reverts a sentence, your raw words go through unchanged. The worst case is always your own words, never nothing and never a rewrite you didn't intend.

To keep it from stalling on a long dictation, Evoglyph gives cleanup a budget that scales with how much you said: a few seconds for a short phrase, up to about thirty for a long passage, with an input ceiling of roughly two thousand tokens. Run past the budget and the raw transcript is what you get.

Every edit is checked

Evoglyph does not simply trust the model's output. Each edit it makes is validated, span by span, by a deterministic safety net that reverts anything doubtful to your raw words. See the span-by-span safety net for how that works.

Where it came from

The current adapter was built clean-room, with no closed-source model in its lineage. Its training labels come from an openly licensed teacher model, Qwen3-235B-A22B (Apache-2.0), and there is an auditable manifest of those labels. The honest framing is that this is a clean-room fine-tune built on openly licensed teachers with an auditable label manifest, not a claim that it is provably free of any closed-source influence.

The base model is openly licensed too, with a caveat worth stating precisely. Liquid AI ships LFM2-2.6B under its own LFM Open License, which grants free commercial use only while the company using it stays under $10 million in annual revenue. That is a condition on Evoglyph's business, not on your dictation; as a user it means nothing for you. The adapter Evoglyph trained on top of that base is its own proprietary work.

How it got better

The adapter has been rebuilt and improved in stages rather than shipped once and frozen. An early clean-room rebuild established the no-closed-source lineage. A later revision pivoted to a lighter, more surgical style of editing. The current version added homophone correction on top of that, with its context-homophone accuracy climbing from roughly half of the cases it should fix to about nine in ten, while its precision stayed near perfect. Earlier adapters stay bundled as fallbacks, so cleanup keeps working even if the newest one is ever unavailable.

Adding homophone correction did not slow cleanup down; its latency stayed essentially flat. For measured, methodology-backed latency across the whole pipeline, see the benchmark page.

The cleanup model sits at the centre of three connected pieces. Its base is Liquid AI's LFM2-2.6B. Homophone correction is a skill trained into this same adapter, and the span-by-span safety net validates everything this adapter produces before it reaches your text. To toggle cleanup on or off and edit the prompt that drives it, see AI cleanup.

The cleanup model