Instructions to use intrect/VELA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use intrect/VELA with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="intrect/VELA")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("intrect/VELA")
model = AutoModelForCausalLM.from_pretrained("intrect/VELA")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

llama-cpp-python

How to use intrect/VELA with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="intrect/VELA",
	filename="vela-dpo-v6-q4_k_m.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

MLX

How to use intrect/VELA with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("intrect/VELA")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use intrect/VELA with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf intrect/VELA:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf intrect/VELA:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf intrect/VELA:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf intrect/VELA: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 intrect/VELA:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf intrect/VELA: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 intrect/VELA:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf intrect/VELA:Q4_K_M

Use Docker

docker model run hf.co/intrect/VELA:Q4_K_M

LM Studio
Jan

vLLM

How to use intrect/VELA with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/intrect/VELA:Q4_K_M

SGLang

How to use intrect/VELA with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "intrect/VELA" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "intrect/VELA",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "intrect/VELA" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "intrect/VELA",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use intrect/VELA with Ollama:
```
ollama run hf.co/intrect/VELA:Q4_K_M
```

Unsloth Studio new

How to use intrect/VELA with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for intrect/VELA to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for intrect/VELA to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for intrect/VELA to start chatting

Pi new

How to use intrect/VELA with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "intrect/VELA"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "intrect/VELA"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use intrect/VELA with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "intrect/VELA"

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 intrect/VELA

Run Hermes

hermes

MLX LM

How to use intrect/VELA with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "intrect/VELA"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "intrect/VELA"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "intrect/VELA",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Docker Model Runner
How to use intrect/VELA with Docker Model Runner:
```
docker model run hf.co/intrect/VELA:Q4_K_M
```

Lemonade

How to use intrect/VELA with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull intrect/VELA:Q4_K_M

Run and chat with the model

lemonade run user.VELA-Q4_K_M

List all available models

lemonade list

intrect commited on Apr 1

Commit

1bc7a0f

verified ·

1 Parent(s): de3e104

docs: 벤치마크 섹션 한국어로 변환

Browse files

Files changed (1) hide show

README.md +53 -54

README.md CHANGED Viewed

@@ -194,77 +194,76 @@ Qwen/Qwen2.5-7B-Instruct
 ## Benchmarks
-### Korean LLM Benchmark (KMMLU + HAE-RAE)
-All models evaluated with **Q4_K_M quantization**, **0-shot**, using `lm-evaluation-harness` v0.4.9 + `llama.cpp` on Apple M1 Max 32GB.
-#### KMMLU (Korean MMLU, 10 subjects)
-| Subject | VELA DPO v6 | Qwen2.5-7B-Instruct | EXAONE-3.5-7.8B |
-|---------|:-----------:|:--------------------:|:---------------:|
-| Marketing | **75.7** | 72.5 | 75.6 |
-| Computer Science | **73.7** | 69.7 | 69.7 |
-| Management | 54.0 | 55.2 | **57.3** |
-| Political Science | 49.0 | 49.3 | **56.0** |
-| Economics | 45.4 | 47.7 | **51.5** |
-| Law | 43.4 | 46.1 | **49.9** |
-| Psychology | 39.2 | 39.3 | **45.7** |
-| Accounting | 38.0 | 33.0 | **42.0** |
-| Math | **33.0** | **33.7** | 27.7 |
-| Korean History | **31.0** | 29.0 | 22.0 |
-| **Average** | **48.2** | **47.6** | **49.7** |
-#### HAE-RAE Bench (Korean-specific)
-| Subtask | VELA DPO v6 | Qwen2.5-7B-Instruct | EXAONE-3.5-7.8B |
-|---------|:-----------:|:--------------------:|:---------------:|
-| Rare Words | 69.9 | 68.4 | **78.8** |
-| Standard Nomenclature | 64.7 | 66.0 | **71.9** |
-| Loan Words | 48.5 | 57.4 | **81.1** |
-| History | 45.7 | 42.6 | **77.7** |
-| General Knowledge | **44.3** | 42.1 | 44.3 |
-| **Average** | **54.5** | **55.3** | **70.7** |
-#### Key Findings
-- **No catastrophic forgetting**: VELA maintains baseline Qwen2.5 capabilities (KMMLU avg 48.2% vs 47.6%) despite domain-specific fine-tuning
-- **Domain transfer**: Finance-related subjects improved — Marketing (+3.2%), Computer Science (+4.0%), Accounting (+5.0%) vs base Qwen2.5
-- **Competitive with Korean-native models**: VELA beats EXAONE-3.5-7.8B (LG AI Research) in 4/10 KMMLU subjects despite EXAONE being pre-trained with large-scale Korean corpus
-### Quantization Benchmark (GGUF)
 RTX 3060 12GB, llama-cpp-python, `n_gpu_layers=-1`, `n_ctx=4096`
-| Format | Speed | Chinese Leak | Quality |
-|--------|-------|--------------|---------|
-| **Q4_K_M (v6)** | **36 tok/s** | 0/5 CLEAN | RT + Report OK |
-> Stress test 5 runs: Synthesis + 3K Reasoning Trace alternating — **zero Chinese leak** in all runs
-### MLX Benchmark (Apple Silicon)
-M1 Max 32GB, MLX 4-bit quantization
-| Config | Quantization | Load Time | Speed | Memory |
-|--------|-------------|-----------|-------|--------|
-| **MLX 4-bit** | 4-bit (4.5 bpw) | 0.59s | **15.93 tok/s** | 4.4 GB |
-| PyTorch (CPU) | BF16 | 0.10s | 4.93 tok/s | 0.3 GB |
-| PyTorch + LoRA (CPU) | BF16 | 1.64s | 4.22 tok/s | 14.1 GB |
 MLX 4-bit vs PyTorch CPU:
-- **3.2x** faster inference (15.93 vs 4.93 tok/s)
-- **73%** smaller model size (4 GB vs 15 GB)
-- **68%** less memory (4.4 vs 14.1 GB)
-### DPO Quality Improvements
-| Metric | Before DPO | After DPO |
-|--------|-----------|-----------|
-| Chinese leak | Frequent | **0/10 CLEAN** |
-| English leak | Occasional | Minimal |
-| RT format compliance | ~80% | **~98%** |
-| Korean fluency | Good | **Excellent** |
 ---
 ## Usage

 ## Benchmarks
+### 한국어 LLM 벤치마크 (KMMLU + HAE-RAE)
+모든 모델 **Q4_K_M 양자화**, **0-shot** 조건으로 평가. `lm-evaluation-harness` v0.4.9 + `llama.cpp`, Apple M1 Max 32GB 환경.
+#### KMMLU (한국어 MMLU, 10과목)
+| 과목 | VELA DPO v6 | Qwen2.5-7B-Instruct | EXAONE-3.5-7.8B |
+|------|:-----------:|:--------------------:|:---------------:|
+| 마케팅 | **75.7** | 72.5 | 75.6 |
+| 컴퓨터과학 | **73.7** | 69.7 | 69.7 |
+| 경영학 | 54.0 | 55.2 | **57.3** |
+| 정치사회학 | 49.0 | 49.3 | **56.0** |
+| 경제학 | 45.4 | 47.7 | **51.5** |
+| 법학 | 43.4 | 46.1 | **49.9** |
+| 심리학 | 39.2 | 39.3 | **45.7** |
+| 회계 | 38.0 | 33.0 | **42.0** |
+| 수학 | **33.0** | **33.7** | 27.7 |
+| 한국사 | **31.0** | 29.0 | 22.0 |
+| **평균** | **48.2** | **47.6** | **49.7** |
+#### HAE-RAE Bench (한국어 특화)
+| 영역 | VELA DPO v6 | Qwen2.5-7B-Instruct | EXAONE-3.5-7.8B |
+|------|:-----------:|:--------------------:|:---------------:|
+| 희귀어 | 69.9 | 68.4 | **78.8** |
+| 표준명칭 | 64.7 | 66.0 | **71.9** |
+| 외래어 | 48.5 | 57.4 | **81.1** |
+| 한국사 | 45.7 | 42.6 | **77.7** |
+| 일반상식 | **44.3** | 42.1 | 44.3 |
+| **평균** | **54.5** | **55.3** | **70.7** |
+#### 주요 발견
+- **Catastrophic forgetting 없음**: 도메인 특화 fine-tuning 후에도 베이스 모델(Qwen2.5) 능력 유지 (KMMLU 평균 48.2% vs 47.6%)
+- **도메인 전이 효과**: 금융 관련 과목에서 베이스 모델 대비 향상 — 마케팅(+3.2%), 컴퓨터과학(+4.0%), 회계(+5.0%)
+- **한국어 네이티브 모델과 경쟁**: 대규모 한국어 코퍼스로 사전학습된 EXAONE-3.5-7.8B (LG AI Research) 대비 KMMLU 10과목 중 4개에서 우위
+### 양자화 벤치마크 (GGUF)
 RTX 3060 12GB, llama-cpp-python, `n_gpu_layers=-1`, `n_ctx=4096`
+| 포맷 | 속도 | 중국어 Leak | 품질 |
+|------|------|-------------|------|
+| **Q4_K_M (v6)** | **36 tok/s** | 0/5 클린 | RT + 리포트 정상 |
+> 스트레스 테스트 5회: Synthesis + 3K Reasoning Trace 교대 — 전 구간 **중국어 leak 제로**
+### MLX 벤치마크 (Apple Silicon)
+M1 Max 32GB, MLX 4-bit 양자화
+| 구성 | 양자화 | 로딩 시간 | 추론 속도 | 메모리 |
+|------|--------|----------|----------|--------|
+| **MLX 4-bit** | 4-bit (4.5 bpw) | 0.59초 | **15.93 tok/s** | 4.4 GB |
+| PyTorch (CPU) | BF16 | 0.10초 | 4.93 tok/s | 0.3 GB |
+| PyTorch + LoRA (CPU) | BF16 | 1.64초 | 4.22 tok/s | 14.1 GB |
 MLX 4-bit vs PyTorch CPU:
+- 추론 속도 **3.2배** (15.93 vs 4.93 tok/s)
+- 모델 크기 **73% 감소** (4 GB vs 15 GB)
+- 메모리 **68% 절약** (4.4 vs 14.1 GB)
+### DPO 학습 품질 개선
+| 지표 | DPO 전 | DPO 후 |
+|------|--------|--------|
+| 중국어 leak | 빈번 | **0/10 클린** |
+| 영어 leak | 간헐적 | 최소화 |
+| RT 형식 준수율 | ~80% | **~98%** |
+| 한국어 유창성 | 양호 | **우수** |
 ---
 ## Usage