Instructions to use osiharris/usuba-checker-adapters with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries
PEFT
How to use osiharris/usuba-checker-adapters with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use osiharris/usuba-checker-adapters with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf osiharris/usuba-checker-adapters:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf osiharris/usuba-checker-adapters:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf osiharris/usuba-checker-adapters:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf osiharris/usuba-checker-adapters:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf osiharris/usuba-checker-adapters:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf osiharris/usuba-checker-adapters:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf osiharris/usuba-checker-adapters:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf osiharris/usuba-checker-adapters:Q4_K_M

Use Docker

docker model run hf.co/osiharris/usuba-checker-adapters:Q4_K_M

LM Studio
Jan

vLLM

How to use osiharris/usuba-checker-adapters with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "osiharris/usuba-checker-adapters"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "osiharris/usuba-checker-adapters",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/osiharris/usuba-checker-adapters:Q4_K_M

Ollama
How to use osiharris/usuba-checker-adapters with Ollama:
```
ollama run hf.co/osiharris/usuba-checker-adapters:Q4_K_M
```
Unsloth Desktop

How to use osiharris/usuba-checker-adapters with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf osiharris/usuba-checker-adapters:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "osiharris/usuba-checker-adapters:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use osiharris/usuba-checker-adapters with Docker Model Runner:
```
docker model run hf.co/osiharris/usuba-checker-adapters:Q4_K_M
```

Lemonade

How to use osiharris/usuba-checker-adapters with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull osiharris/usuba-checker-adapters:Q4_K_M

Run and chat with the model

lemonade run user.usuba-checker-adapters-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use osiharris/usuba-checker-adapters with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf osiharris/usuba-checker-adapters:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default osiharris/usuba-checker-adapters:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use osiharris/usuba-checker-adapters with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf osiharris/usuba-checker-adapters:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "osiharris/usuba-checker-adapters:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Usuba Checker Model v0.1.0 (run 2)

Status: released as v0.1.0 (2026-10-17). An early release, shared as an invitation to help. Advisory only: do not use it as an automated accept/reject gate. The verdict comes from the Usuba Checker Scripts (v0.1.0); this model reads their report and explains. Release notes: GitHub, model-v0.1.0. Project site: usuba.dev.

What is here

path	what
`adapter/`	LoRA adapter (PEFT, r=32, alpha=64, dropout 0.05, all attention/MLP projections), `best` checkpoint of run 2. `adapter_model.safetensors`: 346,294,736 bytes, sha256 `79cb47183b6577aa046ae2c89f961bd5538f3353ab3eee48455b026946242bb2`
`gguf/usuba-9b-usuba-9b-lora-r7-best-Q4_K_M.gguf`	Adapter merged into the base and quantised to Q4_K_M with llama.cpp. 5,629,108,800 bytes, sha256 `7d2b54a24f335aa31d53a47fca8e0848add83bef8246f0f7f831c0e80c9e4f8e`

Training

Base: Qwen/Qwen3.5-9B (Apache-2.0), revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.
Run 2: 221 train / 23 validation rows, 28 steps, lr 1e-4 cosine, seed 1337, scope prompt v4. Final validation loss 0.263.
Task: read an Omarchy (Quickshell/QML) plugin plus the Checker Scripts report; answer in a fixed JSON format with a verdict and findings.

How it was scored — read this first

All scores below are for the bf16 merge of base + adapter, not for the Q4_K_M file in this repo. The Q4_K_M file has not been tested on any exam set; its speed and accuracy are unmeasured. One-shot, same harness, schema, quote filter and greedy decoding as the stock Qwen3.5-9B comparison. Labels are machine drafts; invalid answers count as wrong. Sets are small (Dell 75: 28 broken, 47 working; S4 60: mined upstream bug locations; HANCORE holdout 329: 300 positives, 29 negatives), so differences of a few points are noise. Source: docs/post/draft.md (usuba-checker-scripts) and R7_RUN2_EXAM.md.

What it did well

It learned the fixed review format. Valid-answer rate, Usuba run 2 vs stock Qwen3.5-9B:

exam set	Usuba run 2	stock Qwen3.5-9B
Dell 75	0.987	0.893
S4 60	0.917	0.567
HANCORE holdout 329	0.939	0.638

Against the other open models in the same one-shot panel (Set 2), it has the best valid-answer rate on all three sets (Dell / S4 / holdout):

model	valid: Dell / S4 / holdout
Usuba run 2	0.987 / 0.917 / 0.939
Stock Qwen3.5 9B	0.893 / 0.567 / 0.638
Gemma 4 26B	0.853 / 0.783 / 0.824
Qwen3-Coder 30B	0.840 / 0.733 / 0.763
GLM 4.7 Flash	0.880 / 0.717 / 0.781
Nemotron 3 Nano 30B	0.813 / 0.633 / 0.650
Ministral 14B	0.680 / 0.467 / 0.571
Llama 3.1 8B	0.493 / 0.417 / 0.453

Other differences from stock (also within small-sample noise):

Dell 75: it flags more broken plugins (0.607 vs 0.286).
S4 60: bugs found 8 of 60 vs 4 of 60 for stock (within noise).
It keeps far fewer of its own added findings after the quote filter: 13 vs 47 on S4, 64 vs 139 on the holdout, 30 vs 41 on Dell.
HANCORE holdout: 8 of 29 clean plugins got a blocker or an invalid answer, vs 11 of 29 for stock.
With the Checker feedback loop (Set 3) it stays at 0.93 [0.90-0.95] valid on the holdout; 70 of 75 Dell answers already passed every Checker check before any revision.

Where it did not beat stock

Verdict accuracy is flat vs stock on the Dell 75: 0.480 [0.37-0.59] vs 0.507 [0.40-0.61]; the intervals overlap heavily. With the Checker loop: 0.47 [0.36-0.57] vs 0.57 [0.47-0.68] for stock + loop.
The Checker Scripts alone score 0.773 [0.667-0.853] on the Dell 75 with no model. In one-shot runs no model in the panel is above 0.56 (Gemma 4 26B); the best with the loop is Gemma 4 26B at 0.65 [0.55-0.76]. Always answering "broken" scores 0.373.
False alarms on working Dell plugins rose from 0.277 (stock) to 0.574; on the 47 working Dell plugins it added a blocker 25 times vs 9 for stock. It rarely disagrees with the Checker (dismissed 3 of 80 Checker findings on Dell, vs 43 of 71 for stock).
The valid-answer charts show format reliability, not better judgment.

Limitations

Only 221 training rows; one run, one seed; machine-draft labels; the Q4_K_M file is untested; Dell results come from one laptop (Omarchy 4.0.4).

Reproducing

REPRODUCE.md lists the exact base revision, data hashes, hyperparameters, software versions, llama.cpp commit and Checker Scripts commit, and what is still missing. training/ holds the run metadata, loss log and pinned requirements. The training data (public subset, with a row index) and the evaluation outputs of the latest round are in the dataset repo (run2_data/, eval_2026-10-16/). Verify this repo with SHA256SUMS (sha256sum -c SHA256SUMS; adapter/SHA256SUMS is the instance's own file and also lists an adapter/README.md that is not uploaded).

Licence

The Usuba model weights and adapters (this repo) are Apache-2.0, matching the base model (Qwen/Qwen3.5-9B). Training data keeps each source plugin's licence (MIT, Apache-2.0, ISC); attribution in NOTICE here and sources/licenses/ in the GitHub repo. The Checker Scripts and all code and docs in the GitHub repos are MIT (LICENSE-repo-MIT.txt).

Model tree for osiharris/usuba-checker-adapters

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Adapter

(955)

this model

osiharris
/

usuba-checker-adapters

Usuba Checker Model v0.1.0 (run 2)

What is here

Training

How it was scored — read this first

What it did well

Where it did not beat stock

Limitations

Reproducing

Licence

Links

Model tree for osiharris/usuba-checker-adapters