Modelpublic

Hanish/quill-fix-v1

sourceHugging Faceapache-2.0updated 12d agoView on Hugging Face

0likes61downloads

Model Card

Quill fix-v1

On-device grammar fix / voice-preserving rewrite model for Quill, a private AI keyboard and chat app. Fine-tuned from Qwen/Qwen3.5-0.8B-Base with LoRA (r=32, all projections) on 58.8k instruction rows (CoEdIT fix/rewrite, phone-style typo noising, 12% no-op rows) for 2 epochs. Merged and exported to GGUF with llama.cpp (convert_hf_to_gguf.py --no-nextn; EOS <|im_end|>).

File	Quant	Size	Use
`fix-v1-Q4_K_M.gguf`	Q4KM	529 MB	phones (4 GB+ RAM)
`fix-v1-Q8_0.gguf`	Q8_0	812 MB	higher precision

Prompt format (ChatML)

<|im_start|>system
You fix text typed on a phone. Keep the writer's voice exactly.
Fix only spelling, grammar, and punctuation. Never translate. Never add or remove sentences.
Output the corrected text only.
Voice: {profile}
Never change these words: {lexicon}<|im_end|>
<|im_start|>user
MODE=fix
in: {text}
out:<|im_end|>
<|im_start|>assistant

Modes: fix, rewrite tone={casual|neutral|formal|shorter|longer}.

Held-out eval (150 rows, vs zero-shot base)

exact-match 9% → 43% · GLEU-lite 0.32 → 0.73 · no-op precision 33% → 94% · over-edit rate 43% → 24%.

Example: i dont think were gonna make it on time, traffic is crazy rn → I don't think we're gonna make it on time, traffic is crazy rn.

Requires llama.cpp with the qwen35 architecture (2026-05 or newer).