Downloads · 30 days
50
33% of all-time downloads
the-acai/iki-gemma-cleanup
iki-gemma-cleanup is a machine learning model from the-acai. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as gemma.
Fine-tuned cleanup model for Iki, the on-device macOS dictation app. QLoRA r32/α32, 3 epochs, merged into google/gemma-4-E2B-it and quantized to Q40 with llama.cpp.
Downloads · 30 days
50
33% of all-time downloads
All-time downloads
153
Public
Repo size
5.9 GB
Likes
0
Public
Click a slice to open those files.
.gguf5.9 GB · 100%
From the Hugging Face model README
Fine-tuned cleanup model for Iki, the on-device macOS dictation app.
QLoRA r32/α32, 3 epochs, merged into google/gemma-4-E2B-it and quantized to Q4_0 with llama.cpp.
Training data is synthetic/public only — no user dictation, no PHI: in-repo synthetic clinical/social-work lines,
templated numeral-convention lines, the Iki benchmark references, and a wikitext-103 sample (CC BY-SA), corrupted with
Iki's generator (iki-bench generate-cleanup-dataset, dataset v6, 9,151 rows). See docs/MODEL_PROVENANCE.md in the Iki repo.
Gates vs stock E2B Q4_0 (Iki's harness): cleanup-corpus exact match 54.7% → 73.6%, numeral golden 18/31 → 28/31, profanity preservation 20/20, hand check 50/50 with zero invented content, WER within noise, p95 cleanup latency 1531 → 509 ms.
The Iki app downloads gemma-4-E2B-it-iki6-Q4_0.gguf from this repo on first launch.