Kanso-Koi-0.8

Kanso Koi 0.8 (1.7B)

<!-- builderhelm-header --> <div align="center"> <img src="assets/kanso-koi.png" width="160" alt="Kanso Koi"/> </div> <div align="center"> <a href="https://getzenvoice.com"><img alt="ZenVoice: Homepage" src="https://img.shields.io/badge/ZenVoice-Homepage-536af5?labelColor=555"/></a> <a href="https://getzenvoice.com/download"><img alt="Download: ZenVoice for Mac" src="https://img.shields.io/badge/Download-ZenVoice%20for%20Mac-536af5?labelColor=555&logo=apple&logoColor=white"/></a> <br> <a href="https://huggingface.co/BuilderHelm"><img alt="Hugging Face: BuilderHelm" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-BuilderHelm-ffc107?labelColor=555"/></a> <br> <a href="#license"><img alt="License: Apache 2.0" src="https://img.shields.io/badge/License-Apache%202.0-f5de53?labelColor=555"/></a> <img alt="Runs on: Apple Silicon · MLX" src="https://img.shields.io/badge/Runs%20on-Apple%20Silicon%20%C2%B7%20MLX-8b8b8b?labelColor=555&logo=apple&logoColor=white"/> <img alt="Status: Retired" src="https://img.shields.io/badge/Status-Retired-9e9e9e?labelColor=555"/> </div> <!-- builderhelm-header -->

Names. ZenVoice's formatting models are called Kanso since ZenVoice 1.3.1: Kanso Koi 1.0 was ZenPolish v2.2 Lite, Koi 0.9 was ZenPolish v2.2, Koi 0.8 was ZenPolish v2, Kanso Satori 1.0 was ZenPolish v3.2 4B and Satori 0.9 was ZenPolish v3 4B. The repositories were renamed from zen-polish-…; old links and pinned revisions redirect here. Tables below use the development version numbers (v2, v2.1, v2.2, v3, v3.1, v3.2).

Retired. ZenVoice no longer offers this model; it stays here for earlier builds. Use Kanso-Koi-1.0 or Kanso-Satori-1.0.

ZenVoice's on-device dictation-enhancement model. Takes the raw, unpunctuated text produced by speech-to-text engines and restores everything dictation loses: punctuation, capitalization, filler removal, spoken numbers, dates, and times. Runs fully on-device on Apple Silicon via MLX — nothing leaves the Mac.

Base model: fine-tuned (LoRA) from Qwen3-1.7B, then merged and 4-bit quantized. Qwen3 is the starting point; the formatting behavior below is entirely the fine-tune.

Usage

ChatML prompt, thinking disabled:

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "um yeah so we need to fix this bug by tuesday right check the pull request"},
]
# -> "Yeah, so we need to fix this bug by Tuesday, right? Check the pull request."

The exact system prompt is in the ZenVoice repository (training/common.py).

Evaluation

Measured on ZenVoice's frozen eval sets (training/evaluate.py with the corrected word-only WER alignment and punctuation substitution accounting), greedy decoding, vs. stock Qwen3-1.7B-Instruct under the same prompt.

300-row synthetic eval — Koi 0.8 vs stock base:

Clean real-input bench (120 rows): Koi 0.8 7.6% WER vs stock 11.0%, exact match 7.5% vs 5.0% — v2 stays ahead of stock on clean input. Its WER is higher on clean input than on noisy ASR-style input (7.6% vs 5.0%); identity-mix training in v3 closes that gap.

Scope

• English-focused (v3 will broaden language coverage)
• Formatting and cleanup only — never rewrites meaning, never adds content
• Strengths: fillers, casing, punctuation, common dates/times/percent
• Weak spots: email/URL rewriting, spoken money amounts, homophone repair

Training

Pipeline and recipe are in the ZenVoice repo under training/: noisifier (spoken-form degradation of clean text), synthetic + real ASR pair builders, frozen eval sets, mlx-lm LoRA training, and the gate harness. Weights: 4-bit MLX safetensors, 934 MB.

License

Apache-2.0, inherited from the Qwen3-1.7B base.

Published under BuilderHelm AI. Previously imYChaudhary22/Kanso-Koi-0.8; existing repository links and pinned revisions redirect to this organization repository. Model filenames and weights are unchanged.