scribekey/scribekey-cleanup-fast-GGUF

ScribeKey Fast by ScribeKey — GGUF

The evaluated Q8_0 build of ScribeKey Fast, a 350M-class English speech-to-text cleanup model. It receives a transcript, not audio, and returns cleaned text. It is not a general chat assistant.

Release v1, experiment v7-replay-tgt. This Q8_0 is ScribeKey's on-device cleanup model, replacing Sotto. Direct-helper S25 Ultra timing of this exact file is complete; midrange and product end-to-end performance have not been measured. Two public-input BF16 CPU smoke cases also ran twice with identical outputs (bf16-smoke.json); this is an inference check, not a BF16 quality benchmark.

File	Bytes	SHA-256
scribekey-cleanup-fast-350m-v1-Q8_0.gguf	379,219,904 (361.7 MiB)	bbce6fe5dc52080366ea65768f96e8d6265c45c1925d9a2e3d03de32856295cc

Converted and quantized with llama.cpp 4c1a0af40d88c7fbb3b15c85bf2e8016d1d5b64c, ScribeKey's runtime pin. Only the download filename changed; the evaluated bytes and internal experiment metadata are preserved. Lower-bit files have not been included because they have not received this candidate's quality evaluation.

Run it

Use raw completion mode and the exact task prompt. The evaluated GGUF retains an inherited LFM chat template; do not rely on its automatic chat formatting. prompt-template.jinja is a separate task-specific template for integrations that support overriding the embedded template.

llama-server -m scribekey-cleanup-fast-350m-v1-Q8_0.gguf \
  --host 127.0.0.1 --port 8080 -c 4096 -t 2 --temp 0 --seed 42

import json
import urllib.request

raw = "um can you send the updated agenda before lunch"
control = "[Styling: semi-casual] [Structure: prose] [Context: general]"
prompt = f"<|startoftext|>### Input:\n{control}\n{raw}\n\n### Output:\n"
request = urllib.request.Request(
    "http://127.0.0.1:8080/completion",
    data=json.dumps({
        "prompt": prompt, "n_predict": 256, "temperature": 0,
        "top_k": 1, "top_p": 1.0, "repeat_penalty": 1.0, "seed": 42,
        "stop": ["###", "<|im_end|>"],
    }).encode(),
    headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(request, timeout=60) as response:
    print(json.load(response)["content"].strip())

Use one BOS, no system message, no thinking mode and greedy decoding. Controls are styling (casual, semi-casual, semi-formal, formal), structure (prose, lists) and context (general, email). ScribeKey uses semi-casual/prose by default, semi-formal/email in email apps, and performs list formatting in the app after generation.

Evidence and limitations

On hidden CleanBench v1 (3,366 cases, 3,358 scored), under semi-casual/prose/general at the app boundary, this Q8_0 differs from Sotto v6 by +1.9 percentage points inserted exact (95% −0.1 to +3.9) and −0.001 token edit rate (−0.010 to +0.009). It differs from S1-mini by −1.2 points exact (−3.3 to +1.1). These intervals do not establish superiority or equivalence. The benchmark new-value metric is 4.6%; names, numbers, negations and omissions need checking.

The frequently quoted S25 Ultra long-input p95 of 2,101 ms belongs to v7-replay, not this final candidate. The final-artifact S25 direct-helper recheck measured short p50/p95 155/178 ms and long 1,702/2,718 ms with Parakeet resident; midrange and product end-to-end qualification remain pending. See evaluation.json, examples.json and the full model card for scope and known limits. Keep raw transcripts and a fallback; output is not guaranteed faithful. Lists and email layout can also be wrong.

Terms and attribution

The weights inherit LFM Open License v1.0, including its commercial revenue condition, not a blanket Apache or MIT licence. Retain ADDITIONAL_TERMS.md, LICENSE, NOTICE and MODIFICATIONS.md when redistributing and mark your changes. Organisations at or above the licence's US$10M annual-revenue threshold should obtain separate terms from Liquid AI before commercial use. Read section 5; this summary is not a legal opinion.

ScribeKey Fast continues Sotto by juanquivilla, based on LFM2.5-350M-Base by Liquid AI. NOTICE preserves Sotto's MIT notice and credits the training sources, including CC BY/CC BY-SA data. Those source terms remain distinct; this card does not resolve whether share-alike reaches model weights. Source text/audio and hidden benchmark rows are not distributed. The inherited Sotto data history has not been independently reconstructed. No upstream endorsement is implied. The ScribeKey Fast Model Terms v1.0 require the ScribeKey Fast name and ScribeKey publisher credit in model selectors/information and documentation; modified derivatives must credit Derived from ScribeKey Fast by ScribeKey.

The v1 release is the public v1-rc1-attribution package with updated card wording; weight bytes, terms and prompt/runtime controls are unchanged from v1-rc1. The original RC tags are preserved.