tigerisaac/mushu-1-preview

Mushu 1 Preview

Mushu 1 Preview is a 1.2B-parameter, Dragon-style dictation model. It turns plain speech-to-text transcripts containing spoken punctuation, layout, and capitalization commands into faithfully rendered text with no system prompt.

The model is named after Mushu, the dog pictured above.

GGUF-only preview: this repository contains the final Q4_K_M quantization, split into three llama.cpp-compatible shards. Full-precision weights, adapters, and a Safetensors checkpoint are not included.

What it does

Given a transcript as the user message, Mushu returns only the rendered dictation. It follows a classic command-driven dictation register:

• render spoken punctuation commands such as period, comma, colon, question mark, and open quote;
• render new line and new paragraph as layout;
• use cap to capitalize the next dictated word;
• preserve every other dictated word in the same order;
• preserve fillers, stutters, repeated words, casual forms, grammar errors, run-ons, number words, and apparent correction language;
• avoid adding punctuation, rewriting prose, formatting numbers, resolving corrections, or guessing the speaker's intent.

Mushu is a faithful dictation renderer, not a transcript-cleanup model or a general-purpose chat assistant.

Supported commands

Spoken form	Rendered form
period	.
comma	,
question mark	?
exclamation point, exclamation mark	!
colon	:
semicolon	;
hyphen	-
dash	—
ellipsis, dot dot dot	...
new line, newline	one line break
new paragraph, next paragraph	two line breaks
open paren, open parenthesis	(
close paren, close parenthesis	)
open quote	opening double quote
close quote, unquote	closing double quote
cap	capitalize the next word

Command words remain literal when they are part of the sentence, such as “the trial period,” “a new line of products,” or “the salary cap.”

Stateful editing and application-control commands are not supported. Phrases such as scratch that, delete that, undo that, all caps on, and spell that are preserved as dictated.

Quick start

This preview is distributed only as a pre-quantized GGUF for llama.cpp. Point llama.cpp at the first shard; it discovers the remaining shards automatically.

llama-server -m mushu-1-preview-Q4_K_M-00001-of-00003.gguf \
  --host 127.0.0.1 --port 8080 \
  --ctx-size 2048 --parallel 1 --reasoning off

Send the raw transcript as the only user message, with no system message:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "mushu-1-preview",
    "temperature": 0,
    "messages": [
      {
        "role": "user",
        "content": "bring three things colon bread comma milk comma and tea period"
      }
    ]
  }'

Expected assistant content:

Bring three things: bread, milk, and tea.

Input boundary

Mushu was trained for an application pipeline that removes automatic STT periods, commas, colons, semicolons, and sentence dashes before inference. Their absence is intentional: Mushu adds these marks only when the speaker dictates a supported command.

For behavior closest to training, apply the same preprocessing and pass plain transcript text through the model's embedded chat template. Existing question marks and exclamation marks may be preserved, but automatic punctuation from an ASR system can otherwise blur the command-driven contract.

Do not add a system prompt. Use deterministic decoding and preserve leading and trailing whitespace if terminal line-break commands matter to your integration.

Training

• Base model: LiquidAI/LFM2.5-1.2B-Instruct
• Method: supervised fine-tuning with Unsloth
• Training format: one raw transcript user message followed by one rendered assistant response; no system message
• Corpus: 20,591 command-grounded dictation pairs, assembled into 19,561 training and 1,030 validation examples
• Label target: faithful US English Dragon-style dictation rendering
• Published artifact: final checkpoint 1224 as a three-shard Q4_K_M GGUF only

The labels are deterministic renderings of declared commands. The corpus covers punctuation, layout, capitalization, command-versus-literal disambiguation, repeated commands, unsupported-command preservation, and ordinary dictated text that must not be rewritten.

Limitations

• English and the supported US command vocabulary were the training target.
• cap can damage rare, tokenizer-hostile, or internally capitalized names.
• Some ambiguous command words can be misclassified; in particular, literal uses of colon may be rendered as punctuation.
• Long letter-by-letter sequences joined with repeated hyphen commands can lose or substitute characters.
• The model may occasionally act on an unsupported command-like phrase.
• This is a preview release of one lossy quantization. Behavior may differ from the unavailable full-precision checkpoint.

Review outputs before using the model for names, identifiers, medical text, or other exact-copy-sensitive dictation.

License

This fine-tune is derived from LiquidAI/LFM2.5-1.2B-Instruct and is released under the applicable upstream model terms. Review the base model repository before use or redistribution.

Acknowledgments

Built with Liquid AI's LFM2.5 base model and Unsloth. Named for Mushu, who provided supervision of the non-gradient variety.