bol-cleanup-350m-4bit

bol-cleanup-350m-4bit

The transcript-cleanup model for Bol, the open-source voice loop for Claude Code. Turns raw voice dictation into a clean coding prompt without changing its meaning:

in:  um so like refactor the the auth module in auth dot py and uh run the
     tests but don't touch login dot py
out: Refactor the auth module in auth.py and run the tests, but don't touch
     login.py.

LoRA fine-tune of LFM2.5-350M, fused and quantized to 4-bit for MLX. 195MB, 40-100ms per cleanup on Apple Silicon.

Why it exists

Generic 1B-class models failed Bol's safety bar in live testing: they dropped negated clauses ("but don't touch login.py") and parroted few-shot examples. A 350M model tuned on exactly this task is both faster and safer.

Eval (held-out test set)

Training data

Fully synthetic and reproducible (pipeline in the Bol repo): 150 seed coding prompts, rule-corrupted into spoken form (fillers, stutters, "auth dot py", "dash dash verbose"), plus roundtrip pairs where each seed is spoken by macOS say and transcribed by Parakeet to capture real STT error patterns, plus identity pairs so clean text passes through untouched.

Use

With Bol, in ~/.config/bol/config.toml:

[cleanup]
model = "abhiyan10/bol-cleanup-350m-4bit"

Standalone with mlx-lm:

from mlx_lm import load, generate
model, tok = load("abhiyan10/bol-cleanup-350m-4bit")
prompt = tok.apply_chat_template([
    {"role": "system", "content": "Clean this voice dictation for a coding agent. Output only the cleaned text."},
    {"role": "user", "content": "um fix the the login bug in auth dot py"},
], add_generation_prompt=True)
print(generate(model, tok, prompt=prompt, max_tokens=120))