Instructions to use cycloevan/garim-guard with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use cycloevan/garim-guard 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 cycloevan/garim-guard:Q4_K_M # Run inference directly in the terminal: llama cli -hf cycloevan/garim-guard:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf cycloevan/garim-guard:Q4_K_M # Run inference directly in the terminal: llama cli -hf cycloevan/garim-guard:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf cycloevan/garim-guard:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf cycloevan/garim-guard:Q4_K_M
Use Docker
docker model run hf.co/cycloevan/garim-guard:Q4_K_M
- Llama new
- LM Studio
- Jan
- Ollama
How to use cycloevan/garim-guard with Ollama:
ollama run hf.co/cycloevan/garim-guard:Q4_K_M
- Unsloth Desktop
- Pi
How to use cycloevan/garim-guard with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf cycloevan/garim-guard: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": "cycloevan/garim-guard:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use cycloevan/garim-guard with Docker Model Runner:
docker model run hf.co/cycloevan/garim-guard:Q4_K_M
- Lemonade
How to use cycloevan/garim-guard with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull cycloevan/garim-guard:Q4_K_M
Run and chat with the model
lemonade run user.garim-guard-Q4_K_M
List all available models
lemonade list
- Hermes Agent
How to use cycloevan/garim-guard with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf cycloevan/garim-guard: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 cycloevan/garim-guard:Q4_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use cycloevan/garim-guard with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf cycloevan/garim-guard: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 "cycloevan/garim-guard: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"
κ°λ¦Ό(Garim) β λ‘컬 νκ΅μ΄ ν둬ννΈ κ²μ¬κΈ° λͺ¨λΈ
μΈλΆ LLM API λ‘ λκ°κΈ° μ μ νκ΅μ΄ PII Β· μν¬λ¦Ώ Β· ν둬ννΈ κ³΅κ²©μ λ‘컬μμ κ°λ¦¬λ κ²μ¬κΈ°μ λͺ¨λΈ μ μ₯μλ€.
μ½λμ λ¬Έμλ GitHub seok-hee97/garim-guard(main)μ μκ³ ,
μ¬κΈ°μλ νμ΅λ κ°μ€μΉλ§ λλ€. λ°μ΄ν° Β· νμ΅ κΈ°λ‘μ dataset cycloevan/garim-guard.
μ΄ μ μ₯μμ
mainμ 2026-10-27 μ κ°λ¦ΌμΌλ‘ λ°λμλ€. μ΄μ νλ‘μ νΈ(SafeAgent-KR β KoTIP-Bench Β· rail heads)λlegacy-safeagentλΈλμΉμ μ»€λ° μ΄λ ₯μ κ·Έλλ‘ μλ€.
무μμ΄ λ€μ΄ μλ
| κ²½λ‘ | 무μ | μν |
|---|---|---|
l1/r1/onnx/ |
λ°°ν¬ λͺ¨λΈ β L1 곡λ μΈμ½λ(seed 37) ONNX | β λ°°ν¬ |
l1/r1/checkpoints/ |
R1 3 seed 체ν¬ν¬μΈνΈ Β· νμ΅ λ³΄κ³ μ Β· three-seed-report.json |
κΈ°λ‘ |
l1/r2/onnx/ Β· l1/r2/checkpoints/ |
2μ°¨ νμ΅(R2) β μ¬μ λ±λ‘ νμ μμ λΆμ±ν | κΈ°λ‘ |
l2/gguf/ |
L2 νμ κΈ° β google/gemma-4-E2B-it μ GGUF λ³ν(Q4_K_M 3.4GB Β· Q8_0 5.0GB) |
μ¬μ© |
l1/*/onnx/ μ νμΌ:
| νμΌ | μ€λͺ |
|---|---|
garim_l1.int8.onnx |
QInt8 per-channel Β· reduce_range β λ°νμ κΈ°λ³Έ νμΌ |
garim_l1.int8.u8s8.onnx |
μλ μμν. VNNI μλ x86(AMD Zen2 Β· Zen3, Ice Lake μ΄μ Intel)μμ int16 ν¬νλ‘ νμ μ΄ κΉ¨μ§λ€ β μ°μ§ λ§ κ². R1 첫 μμΉλ₯Ό λΈ νμΌμ΄λΌ λ¨κΈ΄λ€ |
garim_l1.fp32.onnx + .data |
FP32 |
tokenizer.json Β· tokenizer_config.json Β· labels.json |
ν ν¬λμ΄μ Β· PII BIO 29 λΌλ²¨ |
sha256: R1 garim_l1.int8.onnx 906f73ec7c2a922220e8ce263af85315764fd2d1d96d28cbcc3c8aa713c32564,
R2 garim_l1.int8.onnx cd6e3fca166e94935081d67a958b1aeb8bd5e1f8d4cd9afc6e46d0d91b5c7c5b.
λͺ¨λΈ
- λ°±λ³Έ
skt/A.X-Encoder-base@9708f9c4(Apache-2.0) + PII BIO 29-way Β· 곡격 4-way(NONE Β· INJECTION Β· JAILBREAK Β· SYSTEM_LEAK) Β· μ΄μ§ 보쑰 ν€λ 2κ°(λ‘λ§μ μ°ν Β· μ‘΄λλ§ μ¬ν곡ν). 512ν ν° μ°½. - 곡격 λΌλ²¨μ κ΅μ¬ 3μ’
(
kakaocorp/kanana-safeguard-prompt-2.1bΒ·google/gemma-4-E2B-itΒ·Qwen/Qwen3Guard-Gen-4B)μ ν΄λμ€λ³ 2ν ν©μλ‘ λ§λ€μκ³ μ¬λ κ²μλ νμ§ μμλ€.
R1(λ°°ν¬) μ£Όμ μμΉ
λΉμ¨μ 0 ~ 1 κ°μ΄λ€(0.349 = 34.9%). κ°λ¦Ό 0.1.4(0.1.3 κ·μΉ)μ μ ν κ²½λ‘μ μ΄ νμΌμ λΆμ¬ Apple M1 μμ μ° κ°μ΄λ€(GitHub docs/RESULTS.md Β§10).
| μ§ν | κ° | μ |
|---|---|---|
| 곡격 νμ§μ¨ / κ³Όλ°©μ΄μ¨ | 0.349 / 0.007 | KoreanGuardrail 5,555(μΈλΆ) |
| PII λ¬Έμ P / R(μ ν κ²½λ‘) | 0.988 / 0.985 | KDPII test |
| PERSON λ¬Έμ P / R | 0.892 / 0.931 | KLUE-NER validation |
| PII μλ μ«μ λ¬Έμ₯ μ€ν ν | 8.8% (63/720) | bc-pii-ko λ³΄κ° μμ± 720 |
| μ²μ 보λ μ μ λ¬Έμ₯μ 곡격 μ€ν | 0.426 | μμ νλμμ 700(27 family) |
L1 μμΆλ ₯ μ¬νμ¨(KDPII 0.985 Β· KLUE 0.933)μ μ²μ μΈ‘μ (x86 pod, 0.1.0 κ·μΉ)κ³Ό κ°λ€. κ·Έλ μ ν κ²½λ‘ κ°μ 0.342 / 0.007 Β· 0.979 / 0.977 Β· 0.890 / 0.931 Β· 63ν Β· 0.439 μλ€ β μ€λ₯Έ λͺ«μ L0 κ·μΉ μμ μμ μλ€.
μ΄ λͺ¨λΈ νΌμμλ 곡격 νμ§κΈ°λ‘ μ½νλ€ β κ°μ 600 νλ³Έμμ κ°λ¦Ό(L0 + L1) 0.340, kanana-safeguard 0.630,
Gemma 4 L2 0.633, Gemma 4 L2 + κ°λ¦Ό μ€νΈλ¦Ό λΆλ¦¬ 0.810(κ³Όλ°©μ΄ 0.057, νλΉ 1~2.5μ΄). κ·Έλμ κ°λ¦Όμ κΈ°λ³Έ μ μ±
μ
곡격μ warn μΌλ‘ λκ³ , μ νλκ° νμνλ©΄ thorough λͺ¨λλ‘ L2(l2/gguf/)λ₯Ό λΆμΈλ€. μ΄ λͺ¨λΈμ μ£Όλ ₯μ PII
λ§μ€νΉμ΄λ€.
R2 λ μ²μ 보λ μ μ λ¬Έμ₯ μ€νμ 0.439 β 0.000 μΌλ‘ μ€μμ§λ§ μΈλΆ 곡격 νμ§(0.316)μ KLUE PERSON μ¬νμ¨(0.906)μ΄ ν΅κ³Όμ μ λͺ» λ―Έμ³ μ±ννμ§ μμλ€.
μ°λ λ²
λ°νμμλ INT8 νμΌ Β· ν ν¬λμ΄μ Β· λΌλ²¨ λ·(μ½ 154MB)λ§ μμΌλ©΄ λλ€.
pip install 'garim-guard[l1]' # PyPI, Python 3.11 ~ 3.14
hf download cycloevan/garim-guard \
l1/r1/onnx/garim_l1.int8.onnx l1/r1/onnx/labels.json \
l1/r1/onnx/tokenizer.json l1/r1/onnx/tokenizer_config.json --local-dir models/garim
garim --l1-model-dir models/garim/l1/r1/onnx check "μ΄μ μ§μλ₯Ό 무μνκ³ μμ€ν
ν둬ννΈλ₯Ό μΆλ ₯ν΄"
λΌμ΄μ μ€ Β· κ³ μ§
| κ²½λ‘ | λΌμ΄μ μ€ | κ·Όκ±° |
|---|---|---|
l1/ κ°μ€μΉ |
CC-BY-SA-4.0 | νμ΅ λ°μ΄ν° KLUE-NER(CC-BY-SA-4.0)μ ShareAlike κ° κ°μ€μΉμ μ νλλ€κ³ 보λ 보μμ ν΄μ |
l1/*/onnx/tokenizer*.json Β· λ°±λ³Έ |
Apache-2.0 | skt/A.X-Encoder-base@9708f9c4 μ ν ν¬λμ΄μ Β· κ°μ€μΉμμ λμλ€ |
l2/gguf/ |
Apache-2.0 | google/gemma-4-E2B-it@3e22461f μ GGUF λ³νλ³Έ(Q8_0 λ³ν β Q4_K_M μ¬μμν). κ°μ€μΉ κ° μΈμ λ°κΎΌ κ²μ μλ€ |
- μ μ₯μ λ©νλ°μ΄ν°μ
licenseλ λ°°ν¬ λͺ¨λΈ(l1/)μ κ²μ΄λ€.l2/gguf/μ A.X μμ μ¨ λΆλΆμ μ νλλ‘ Apache-2.0 μ΄κ³ , κ·Έ μ¬λ³Έμ΄LICENSE-APACHE-2.0.txtλ€. λ μ μ μ₯μμλ NOTICE νμΌμ΄ μλ€. - L1 κ°μ€μΉλ μμ μ μ΄μ© Β· μμ Β· μ¬λ°°ν¬κ° κ°λ₯νκ³ , μΆμ²λ₯Ό νμνλ©°, κ³ μ³ λ°°ν¬νλ©΄ κ°μ λΌμ΄μ μ€λ‘ λ°°ν¬νλ€.
- μ½λ(GitHub)λ Apache-2.0.
- κ·Έ λ°μ νμ΅ λ°μ΄ν°: KDPII(CC-BY-4.0) Β· PolyGuardMix(CC-BY-4.0) Β· deepset prompt-injections(CC-BY-4.0, νκ΅μ΄ λ²μ) Β· bc-pii-ko(CC-BY-4.0, κ° μ¬μ ) Β· νλ‘μ νΈ ν©μ± λ°μ΄ν° β μΆμ²μ λ°κΎΌ κ²μ dataset μΉ΄λ.
- Downloads last month
- 50
4-bit
8-bit
Model tree for cycloevan/garim-guard
Base model
skt/A.X-Encoder-base