Usuba Checker Model v0.1.0

The first release of the Usuba Checker Model: a small local model that reads an Omarchy plugin together with the Usuba Checker Scripts report and explains what to fix. v0.1.0 is Usuba run 2, trained on October 16, 2026. Project site: usuba.dev.

Download (on Hugging Face, pinned to the release commit)

File	Size	sha256
4-bit GGUF for llama.cpp (Q4_K_M)	5.63 GB	7d2b54a24f335aa31d53a47fca8e0848add83bef8246f0f7f831c0e80c9e4f8e
LoRA adapter for Qwen/Qwen3.5-9B	about 350 MB	79cb47183b6577aa046ae2c89f961bd5538f3353ab3eee48455b026946242bb2

llama-server -m usuba-9b-usuba-9b-lora-r7-best-Q4_K_M.gguf -c 32768

The training data, exam outputs and REPRODUCE.md are in osiharris/usuba-checker-data.

How it did, against its stock base (preliminary results, October 2026; machine-drafted labels, one shot)

• Better: valid answers on every exam (Dell 75: 0.99 vs 0.89; S4 60: 0.92 vs 0.57; holdout 329: 0.94 vs 0.64), and broken plugins flagged (0.61 vs 0.29).
• On par: Dell 75 verdict accuracy (0.48 vs 0.51) and bugs found (S4 60).
• Worse: false alarms on working plugins (0.57 vs 0.28).

The Checker Scripts alone score 0.773 on the Dell 75, so their verdict is the one to trust. Use the model as an advisor.

Known gaps:

• The scores are for the bf16 merge; the 4-bit file hasn't been scored yet.
• One run, one seed.
• The training-code commit is unverified.
• There's no one-command reviewer for your own plugin yet.

The full write-up: How Usuba was built and tested.

Licence: weights and adapter Apache-2.0, matching Qwen3.5-9B; code MIT, © 2026 OutboundSync.