Instructions to use wish418/carbix-v3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use wish418/carbix-v3 with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use wish418/carbix-v3 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 wish418/carbix-v3 # Run inference directly in the terminal: llama cli -hf wish418/carbix-v3
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf wish418/carbix-v3 # Run inference directly in the terminal: llama cli -hf wish418/carbix-v3
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 wish418/carbix-v3 # Run inference directly in the terminal: ./llama-cli -hf wish418/carbix-v3
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 wish418/carbix-v3 # Run inference directly in the terminal: ./build/bin/llama-cli -hf wish418/carbix-v3
Use Docker
docker model run hf.co/wish418/carbix-v3
- LM Studio
- Jan
- vLLM
How to use wish418/carbix-v3 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "wish418/carbix-v3" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "wish418/carbix-v3", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/wish418/carbix-v3
- Ollama
How to use wish418/carbix-v3 with Ollama:
ollama run hf.co/wish418/carbix-v3
- Unsloth Desktop
- Pi
How to use wish418/carbix-v3 with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf wish418/carbix-v3
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": "wish418/carbix-v3" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use wish418/carbix-v3 with Docker Model Runner:
docker model run hf.co/wish418/carbix-v3
- Lemonade
How to use wish418/carbix-v3 with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull wish418/carbix-v3
Run and chat with the model
lemonade run user.carbix-v3-{{QUANT_TAG}}List all available models
lemonade list
- Hermes Agent
How to use wish418/carbix-v3 with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf wish418/carbix-v3
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 wish418/carbix-v3
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use wish418/carbix-v3 with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf wish418/carbix-v3
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 "wish418/carbix-v3" \ --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"
carbix-v3
ํ๊ตญ์ด ๊ณต๊ณผ๊ธ ๊ณ ์ง์(์ ๊ธฐยท๊ฐ์คยท์ด์ก)์์ ๊ตฌ์กฐํ ํ๋๋ฅผ ๋ฝ์๋ด๋ VLM.
Qwen/Qwen3-VL-4B-Instruct ๋ฅผ LoRA SFT ๋ก ํ์ธํ๋ํ ๋ค ๋ณํฉํ๊ณ GGUF q8_0 ์ผ๋ก
์์ํํ๋ค. carbix-engine ์ ๋ก์ปฌ ์ถ๋ก ๋ฐฑ์๋๋ก ์ด๋ค.
ํ์ผ
| ํ์ผ | ํฌ๊ธฐ | ์ฉ๋ |
|---|---|---|
carbix-v3-q8.gguf |
4.28 GB | ์๋น ๊ฐ์ค์น (llama.cpp) |
carbix-v3-mmproj.gguf |
836 MB | ๋น์ ์ธ์ฝ๋(๋ฉํฐ๋ชจ๋ฌ ํ๋ก์ ํฐ) โ ์์ผ๋ฉด ์ด๋ฏธ์ง ์ ๋ ฅ ๋ถ๊ฐ |
lora_adapter_v3/ |
144 MB | PEFT ์ด๋ํฐ ์๋ณธ. ์ฌ๋ณํฉยท์ฌํ์ต์ฉ |
test.jsonl |
2.4 MB | ํ๊ฐ์
657๊ฑด + ํ์ต์ ์ค์ ๋ก ์ด ํ๋กฌํํธ (--prompt-from ์ฉ) |
lora_adapter_v3/ ๋ ์ต์ข
์ฐ์ถ๋ฌผ(adapter_model.safetensors)๋ง ๋ด์๋ค. ํ์ต ์ค๊ฐ
์ฒดํฌํฌ์ธํธ 3๊ฐ(checkpoint-2750/3000/3125, ๊ฐ 379MB, ์ตํฐ๋ง์ด์ ์ํ ํฌํจ)๋
์ฉ๋ ๋๋น ์ค์ต์ด ์์ด ์ ์ธํ๋ค. ์ต์ข
์คํ
์ checkpoint-3125 ๋ก ์ด 3,125 ์คํ
์ด๋ค.
์คํ
llama-server \
-m carbix-v3-q8.gguf \
--mmproj carbix-v3-mmproj.gguf \
--host 127.0.0.1 --port 8001 \
-ngl 99 -c 8192 --parallel 2 \
--image-min-tokens 1024 \
--jinja
OpenAI ํธํ API ๋ก ๋ฌ๋ค. carbix-engine ์์๋:
python eval/predict_one.py --image bill.jpg --category gas \
--base-url http://127.0.0.1:8001 --prompt-from test.jsonl
-ngl 99์ ์ฒด ๋ ์ด์ด GPU ์คํ๋ก๋. q8_0 + mmproj ๋ L4 24GB ์ ๋ณ๋ ฌ ์ฌ๋กฏ์ ์ฌ๋ฌ ๊ฐ ๋์๋ ์ฌ์ ๋กญ๋ค.--image-min-tokens 1024Qwen-VL ๊ณ์ด์ ์ด๋ฏธ์ง ํ ํฐ์ด 1024 ๋ฏธ๋ง์ด๋ฉด ์ ํ๋๊ฐ ๋จ์ด์ง๋ค๋ llama.cpp ์์ฒด ๊ฒฝ๊ณ ๋ฅผ ๋ฐ์ํ ๊ฐ.--jinja๋ณํฉ ๋ชจ๋ธ์chat_template.jinja๋ฅผ ์ด๋ค.
ํ์ต
| ํญ๋ชฉ | ๊ฐ |
|---|---|
| ๋ฒ ์ด์ค | Qwen/Qwen3-VL-4B-Instruct |
| ๋ฐฉ๋ฒ | LoRA SFT (transformers + peft ์์ , train/train_hf.py) |
| LoRA | r=16, alpha=32, dropout=0.05, bias=none |
| target_modules | q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj |
| ํ์ต ํ๋ผ๋ฏธํฐ | 33,030,144 / 4,470,845,952 (0.74%) |
| ์ตํฐ๋ง์ด์ | adamw_torch, lr 2e-4, cosine, warmup 5%, weight_decay 0.01 |
| ๋ฐฐ์น | 1 ร grad_accum 4 = effective 4 |
| ์ํญ | 1 (12,500๊ฑด โ 3,125 ์คํ ) |
| ์ ๋ฐ๋ | bf16, gradient checkpointing, ๋น์ ํ์ ๋๊ฒฐ ์ ํจ |
| ํ๋์จ์ด | GCP L4 24GB (g2-standard-4) |
| ์ด๋ฏธ์ง ํด์๋ | shortest_edge=784, longest_edge=1003520 |
| peft | 0.20.0 |
LoRA ๊ฐ(r/alpha/dropout/target_modules)์ ๋๋ดํ lora_adapter_v3/adapter_config.json
์ค์ธก์ด๋ค. carbix-engine ์ train/train_hf.py ๋
์คํธ๋ง์๋ alpha=16 ์ด๋ผ๊ณ ์ ํ
์์ผ๋ ๋ก์ ๊ธฐ์ ์ด๋ค โ argparse ๊ธฐ๋ณธ๊ฐ๊ณผ ์ค์ ํ์ต ๋ก๊ทธ ๋ชจ๋ alpha=32 ๋ค.
๋ฐ์ดํฐ
carbix-augmentation ํ์ดํ๋ผ์ธ์ด ์์ฑํ ํฉ์ฑ ๊ณ ์ง์. ์ค์ ๊ณ ๊ฐ ๋ฌธ์๋ ์ฐ์ง ์์๋ค.
| ํ์ต 12,500 | ๊ฒ์ฆ 1,388 (์ค์ ์ฌ์ฉ 300) | |
|---|---|---|
| ๊ฐ์ค | 10,681 | 1,207 |
| ์ด์ก | 916 | 84 |
| ์ ๊ธฐ | 903 | 97 |
์ฑ๋ฅ
ํฉ์ฑ ํ
์คํธ์
(657๊ฑด, test.jsonl)
| ๋ชจ๋ธ | overall | ํ๊ท ์ง์ฐ | ์๋ฌ |
|---|---|---|---|
| carbix-v3 GGUF q8_0 (llama.cpp) | 99.43% | 3.23s | 0 |
| carbix-v3 ์ด๋ํฐ ์ง์ ๋ก๋ (bf16, transformers) | 98.68% | 11.95s | 0 |
์ฌํ
# ์ด๋ํฐ โ ๋ณํฉ ๋ชจ๋ธ
CARBIX_MODEL_DIR=~/Projects/carbix-model python train/merge_lora.py
# ๋ณํฉ ๋ชจ๋ธ โ GGUF + mmproj
python llama.cpp/convert_hf_to_gguf.py <merged_model> --outtype q8_0
๋ผ์ด์ ์ค
๋ฒ ์ด์ค ๋ชจ๋ธ Qwen/Qwen3-VL-4B-Instruct ์ ๋ผ์ด์ ์ค๋ฅผ ๋ฐ๋ฅธ๋ค. (๋ฐฐํฌ ์ ํ์ธ ํ์)
- Downloads last month
- 114
We're not able to determine the quantization variants.
Model tree for wish418/carbix-v3
Base model
Qwen/Qwen3-VL-4B-Instruct