docs: update model card with GGUF formats, benchmarks, usage examples
Browse files
README.md
CHANGED
|
@@ -10,6 +10,8 @@ tags:
|
|
| 10 |
- stock-analysis
|
| 11 |
- reasoning
|
| 12 |
- dpo
|
|
|
|
|
|
|
| 13 |
base_model: Qwen/Qwen2.5-7B-Instruct
|
| 14 |
pipeline_tag: text-generation
|
| 15 |
---
|
|
@@ -25,51 +27,92 @@ VELA๋ ํ๊ตญ ์ฃผ์์์ฅ ๋ด์ค ๋ถ์ ๋ฐ ํฌ์ ๋ฆฌ์์น๋ฅผ ์ํด ํนํ
|
|
| 25 |
| ํญ๋ชฉ | ๋ด์ฉ |
|
| 26 |
|------|------|
|
| 27 |
| **Base Model** | Qwen/Qwen2.5-7B-Instruct |
|
| 28 |
-
| **Training
|
| 29 |
| **Parameters** | 7.6B |
|
| 30 |
| **Context Length** | 8,192 tokens |
|
| 31 |
-
| **Precision** | BFloat16 |
|
| 32 |
| **License** | Apache 2.0 |
|
| 33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
## Training Pipeline
|
| 35 |
|
| 36 |
```
|
| 37 |
Qwen2.5-7B-Instruct
|
| 38 |
โ
|
| 39 |
SFT (930K samples)
|
| 40 |
-
- ํ๊ตญ ์ฃผ์ ๋ด์ค ๋ถ์
|
| 41 |
-
- ๋ฆฌ์์น ๋ฆฌํฌํธ ์์ฑ
|
| 42 |
-
- Reasoning Trace ํ์ต
|
| 43 |
โ
|
| 44 |
-
DPO
|
| 45 |
- ์ค๊ตญ์ด/์์ด leak ๊ต์
|
| 46 |
- ํ๊ตญ์ด ์ถ๋ ฅ ๊ฐํ
|
| 47 |
- ํ์ ์ค์ ํฅ์
|
| 48 |
โ
|
| 49 |
-
|
| 50 |
```
|
| 51 |
|
| 52 |
## Capabilities
|
| 53 |
|
| 54 |
- **๋ด์ค ์ํฅ ๋ถ์**: ์ฃผ์ ๊ด๋ จ ๋ด์ค์ ์์ฅ ์ํฅ๋ ์์ธก
|
| 55 |
-
- **๋ฆฌ์์น ๋ฆฌํฌํธ ์์ฑ**: ๊ตฌ์กฐํ๋ ํฌ์ ๋ถ์ ๋ณด๊ณ ์
|
| 56 |
-
- **Reasoning Trace**: ๋จ๊ณ๋ณ ๋ถ์ ์ฌ๊ณ ๊ณผ์
|
| 57 |
- **๋ค์ค ์์ค ์ข
ํฉ**: ๋ด์ค, ์์ธ, ์๊ธ ๋ฐ์ดํฐ ํตํฉ ๋ถ์
|
| 58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
## Usage
|
| 60 |
|
| 61 |
-
###
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
|
| 63 |
```python
|
| 64 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 65 |
import torch
|
| 66 |
|
| 67 |
model = AutoModelForCausalLM.from_pretrained(
|
| 68 |
-
"intrect/
|
| 69 |
torch_dtype=torch.bfloat16,
|
| 70 |
device_map="auto"
|
| 71 |
)
|
| 72 |
-
tokenizer = AutoTokenizer.from_pretrained("intrect/
|
| 73 |
|
| 74 |
messages = [
|
| 75 |
{"role": "system", "content": "๋น์ ์ ํ๊ตญ ์ฃผ์ ์ ๋ฌธ ์ ๋๋ฆฌ์คํธ์
๋๋ค."},
|
|
@@ -88,46 +131,66 @@ outputs = model.generate(
|
|
| 88 |
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
| 89 |
```
|
| 90 |
|
| 91 |
-
### vLLM
|
| 92 |
|
| 93 |
```python
|
| 94 |
from vllm import LLM, SamplingParams
|
| 95 |
|
| 96 |
-
llm = LLM(model="intrect/
|
| 97 |
params = SamplingParams(temperature=0.7, max_tokens=1024)
|
| 98 |
|
| 99 |
prompts = ["์ผ์ฑ์ ์ HBM ์์ฅ ์ ๋ง์ ๋ถ์ํด์ฃผ์ธ์."]
|
| 100 |
outputs = llm.generate(prompts, params)
|
| 101 |
```
|
| 102 |
|
| 103 |
-
###
|
| 104 |
-
|
| 105 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
|
| 107 |
## Output Format
|
| 108 |
|
| 109 |
-
VELA๋
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
|
| 111 |
```markdown
|
|
|
|
|
|
|
| 112 |
## Executive Summary
|
| 113 |
[2-3๋ฌธ์ฅ ํต์ฌ ์์ฝ]
|
| 114 |
|
| 115 |
## Key Metrics
|
| 116 |
| ์งํ | ์์น |
|
| 117 |
|------|------|
|
| 118 |
-
| ํ์ฌ๊ฐ | โฉXX,XXX |
|
| 119 |
-
| PER | XX.X |
|
| 120 |
-
| ... | ... |
|
| 121 |
|
| 122 |
## ์์ฅ ๋ํฅ ๋ถ์
|
| 123 |
-
|
| 124 |
-
|
| 125 |
## ๋ฆฌ์คํฌ ์์ธ
|
| 126 |
-
- ๋ฆฌ์คํฌ 1
|
| 127 |
-
- ๋ฆฌ์คํฌ 2
|
| 128 |
-
|
| 129 |
## ํฌ์ ์๊ฒฌ
|
| 130 |
-
[์ข
ํฉ ์๊ฒฌ]
|
| 131 |
```
|
| 132 |
|
| 133 |
## Training Data
|
|
@@ -139,13 +202,11 @@ VELA๋ ๋ค์๊ณผ ๊ฐ์ ๊ตฌ์กฐํ๋ ์ถ๋ ฅ์ ์์ฑํฉ๋๋ค:
|
|
| 139 |
| Reasoning Traces | 5K | ์ฌ๊ณ ๊ณผ์ ํ์ต |
|
| 140 |
| DPO Pairs | 7.7K | ์ ํธ๋ ์ ๋ ฌ |
|
| 141 |
|
| 142 |
-
## DPO
|
| 143 |
-
|
| 144 |
-
DPO v4๋ ๋ค์ ๋ฌธ์ ๋ค์ ํด๊ฒฐํฉ๋๋ค:
|
| 145 |
|
| 146 |
-
- โ
**์ค๊ตญ์ด leak ์ ๊ฑฐ**:
|
| 147 |
- โ
**์์ด leak ๊ฐ์**: ๋ถํ์ํ ์์ด ์ฌ์ฉ ์ต์ํ
|
| 148 |
-
- โ
**ํ์ ์ค์**:
|
| 149 |
- โ
**ํ๊ตญ์ด ํ์ง**: ์์ฐ์ค๋ฌ์ด ํ๊ตญ์ด ํํ
|
| 150 |
|
| 151 |
## Limitations
|
|
@@ -153,6 +214,7 @@ DPO v4๋ ๋ค์ ๋ฌธ์ ๋ค์ ํด๊ฒฐํฉ๋๋ค:
|
|
| 153 |
- ์ค์๊ฐ ์์ธ ๋ฐ์ดํฐ ์ ๊ทผ ๋ถ๊ฐ (์ธ๋ถ API ํ์)
|
| 154 |
- ํฌ์ ์กฐ์ธ์ด ์๋ ์ ๋ณด ์ ๊ณต ๋ชฉ์
|
| 155 |
- 8K ์ปจํ
์คํธ ์ ํ์ผ๋ก ๊ธด ๋ฌธ์ ์ฒ๋ฆฌ ํ๊ณ
|
|
|
|
| 156 |
|
| 157 |
## Citation
|
| 158 |
|
|
@@ -162,7 +224,7 @@ DPO v4๋ ๋ค์ ๋ฌธ์ ๋ค์ ํด๊ฒฐํฉ๋๋ค:
|
|
| 162 |
author={intrect},
|
| 163 |
year={2026},
|
| 164 |
publisher={Hugging Face},
|
| 165 |
-
url={https://huggingface.co/intrect/
|
| 166 |
}
|
| 167 |
```
|
| 168 |
|
|
@@ -170,8 +232,9 @@ DPO v4๋ ๋ค์ ๋ฌธ์ ๋ค์ ํด๊ฒฐํฉ๋๋ค:
|
|
| 170 |
|
| 171 |
| ๋ฒ์ | ๋ ์ง | ๋ณ๊ฒฝ์ฌํญ |
|
| 172 |
|------|------|----------|
|
| 173 |
-
| v1.
|
| 174 |
-
|
|
|
|
|
| 175 |
|
| 176 |
---
|
| 177 |
|
|
|
|
| 10 |
- stock-analysis
|
| 11 |
- reasoning
|
| 12 |
- dpo
|
| 13 |
+
- gguf
|
| 14 |
+
- llama-cpp
|
| 15 |
base_model: Qwen/Qwen2.5-7B-Instruct
|
| 16 |
pipeline_tag: text-generation
|
| 17 |
---
|
|
|
|
| 27 |
| ํญ๋ชฉ | ๋ด์ฉ |
|
| 28 |
|------|------|
|
| 29 |
| **Base Model** | Qwen/Qwen2.5-7B-Instruct |
|
| 30 |
+
| **Training** | SFT (930K) + DPO (7,681 pairs) |
|
| 31 |
| **Parameters** | 7.6B |
|
| 32 |
| **Context Length** | 8,192 tokens |
|
|
|
|
| 33 |
| **License** | Apache 2.0 |
|
| 34 |
|
| 35 |
+
### Available Formats
|
| 36 |
+
|
| 37 |
+
| Format | File | Size | Use Case |
|
| 38 |
+
|--------|------|------|----------|
|
| 39 |
+
| **BF16** (safetensors) | `model.safetensors` | 15 GB | Full precision, GPU inference |
|
| 40 |
+
| **GGUF Q8_0** | `vela-q8_0.gguf` | 7.6 GB | High quality quantized, GPU/CPU |
|
| 41 |
+
| **GGUF Q4_K_M** | `vela-q4_k_m.gguf` | 4.4 GB | Fast & lightweight, GPU/CPU |
|
| 42 |
+
|
| 43 |
## Training Pipeline
|
| 44 |
|
| 45 |
```
|
| 46 |
Qwen2.5-7B-Instruct
|
| 47 |
โ
|
| 48 |
SFT (930K samples)
|
| 49 |
+
- ํ๊ตญ ์ฃผ์ ๋ด์ค ๋ถ์ (412K)
|
| 50 |
+
- ๋ฆฌ์์น ๋ฆฌํฌํธ ์์ฑ (50K)
|
| 51 |
+
- Reasoning Trace ํ์ต (5K)
|
| 52 |
โ
|
| 53 |
+
DPO (7,681 pairs)
|
| 54 |
- ์ค๊ตญ์ด/์์ด leak ๊ต์
|
| 55 |
- ํ๊ตญ์ด ์ถ๋ ฅ ๊ฐํ
|
| 56 |
- ํ์ ์ค์ ํฅ์
|
| 57 |
โ
|
| 58 |
+
VELA
|
| 59 |
```
|
| 60 |
|
| 61 |
## Capabilities
|
| 62 |
|
| 63 |
- **๋ด์ค ์ํฅ ๋ถ์**: ์ฃผ์ ๊ด๋ จ ๋ด์ค์ ์์ฅ ์ํฅ๋ ์์ธก
|
| 64 |
+
- **๋ฆฌ์์น ๋ฆฌํฌํธ ์์ฑ**: ๊ตฌ์กฐํ๋ ํฌ์ ๋ถ์ ๋ณด๊ณ ์ (7๊ฐ ์น์
)
|
| 65 |
+
- **Reasoning Trace**: ๋จ๊ณ๋ณ ๋ถ์ ์ฌ๊ณ ๊ณผ์ (JSON ํ์)
|
| 66 |
- **๋ค์ค ์์ค ์ข
ํฉ**: ๋ด์ค, ์์ธ, ์๊ธ ๋ฐ์ดํฐ ํตํฉ ๋ถ์
|
| 67 |
|
| 68 |
+
## Quantization Benchmark
|
| 69 |
+
|
| 70 |
+
RTX 3060 12GB, llama-cpp-python, n_gpu_layers=-1, n_ctx=4096
|
| 71 |
+
|
| 72 |
+
| Format | Speed (tok/s) | Chinese Leak | Quality |
|
| 73 |
+
|--------|--------------|--------------|---------|
|
| 74 |
+
| **Q4_K_M** | **36 tok/s** | 0/5 CLEAN | Reasoning Trace + Report OK |
|
| 75 |
+
| **Q8_0** | 25 tok/s | 0/5 CLEAN | Reasoning Trace + Report OK |
|
| 76 |
+
|
| 77 |
+
> Stress test: 5ํ ์ฐ์ (Synthesis + 3K Reasoning Trace ๊ต๋) - ์์ชฝ ๋ชจ๋ Chinese leak ์ ๋ก
|
| 78 |
+
|
| 79 |
## Usage
|
| 80 |
|
| 81 |
+
### llama-cpp-python (Recommended for GGUF)
|
| 82 |
+
|
| 83 |
+
```python
|
| 84 |
+
from llama_cpp import Llama
|
| 85 |
+
|
| 86 |
+
model = Llama(
|
| 87 |
+
model_path="vela-q4_k_m.gguf", # or vela-q8_0.gguf
|
| 88 |
+
n_ctx=4096,
|
| 89 |
+
n_gpu_layers=-1, # Full GPU offload
|
| 90 |
+
chat_format="chatml",
|
| 91 |
+
)
|
| 92 |
+
|
| 93 |
+
response = model.create_chat_completion(
|
| 94 |
+
messages=[
|
| 95 |
+
{"role": "system", "content": "๋น์ ์ ํ๊ตญ ์ฃผ์ ์ ๋ฌธ ์ ๋๋ฆฌ์คํธ์
๋๋ค."},
|
| 96 |
+
{"role": "user", "content": "์ผ์ฑ์ ์ HBM ์ฌ์
์ ๋ง์ ๋ถ์ํด์ฃผ์ธ์."},
|
| 97 |
+
],
|
| 98 |
+
max_tokens=1024,
|
| 99 |
+
temperature=0.7,
|
| 100 |
+
)
|
| 101 |
+
print(response["choices"][0]["message"]["content"])
|
| 102 |
+
```
|
| 103 |
+
|
| 104 |
+
### Transformers (BF16)
|
| 105 |
|
| 106 |
```python
|
| 107 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 108 |
import torch
|
| 109 |
|
| 110 |
model = AutoModelForCausalLM.from_pretrained(
|
| 111 |
+
"intrect/VELA",
|
| 112 |
torch_dtype=torch.bfloat16,
|
| 113 |
device_map="auto"
|
| 114 |
)
|
| 115 |
+
tokenizer = AutoTokenizer.from_pretrained("intrect/VELA")
|
| 116 |
|
| 117 |
messages = [
|
| 118 |
{"role": "system", "content": "๋น์ ์ ํ๊ตญ ์ฃผ์ ์ ๋ฌธ ์ ๋๋ฆฌ์คํธ์
๋๋ค."},
|
|
|
|
| 131 |
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
| 132 |
```
|
| 133 |
|
| 134 |
+
### vLLM
|
| 135 |
|
| 136 |
```python
|
| 137 |
from vllm import LLM, SamplingParams
|
| 138 |
|
| 139 |
+
llm = LLM(model="intrect/VELA", dtype="bfloat16")
|
| 140 |
params = SamplingParams(temperature=0.7, max_tokens=1024)
|
| 141 |
|
| 142 |
prompts = ["์ผ์ฑ์ ์ HBM ์์ฅ ์ ๋ง์ ๋ถ์ํด์ฃผ์ธ์."]
|
| 143 |
outputs = llm.generate(prompts, params)
|
| 144 |
```
|
| 145 |
|
| 146 |
+
### Ollama
|
| 147 |
+
|
| 148 |
+
```bash
|
| 149 |
+
# Modelfile
|
| 150 |
+
FROM ./vela-q4_k_m.gguf
|
| 151 |
+
TEMPLATE """<|im_start|>system
|
| 152 |
+
{{ .System }}<|im_end|>
|
| 153 |
+
<|im_start|>user
|
| 154 |
+
{{ .Prompt }}<|im_end|>
|
| 155 |
+
<|im_start|>assistant
|
| 156 |
+
"""
|
| 157 |
+
PARAMETER temperature 0.7
|
| 158 |
+
PARAMETER num_ctx 4096
|
| 159 |
+
```
|
| 160 |
|
| 161 |
## Output Format
|
| 162 |
|
| 163 |
+
VELA๋ ๋ ๊ฐ์ง ์ถ๋ ฅ ๋ชจ๋๋ฅผ ์ง์ํฉ๋๋ค:
|
| 164 |
+
|
| 165 |
+
### 1. Reasoning Trace (๋ถ์ ๊ณผ์ )
|
| 166 |
+
|
| 167 |
+
```json
|
| 168 |
+
{
|
| 169 |
+
"step": 1,
|
| 170 |
+
"thought": "์ผ์ฑ์ ์ HBM3E 12๋จ ์์ฐ ๊ด๋ จ ๋ด์ค ํ์ธ. ์ถ๊ฐ ์์ฃผ ํํฉ๊ณผ ์์ฅ ์ ์ ์จ ํ์
ํ์.",
|
| 171 |
+
"action": "search",
|
| 172 |
+
"query": "์ผ์ฑ์ ์ HBM3E 12๋จ ์์ฃผ ์์ฅ์ ์ ์จ",
|
| 173 |
+
"confidence": 0.45
|
| 174 |
+
}
|
| 175 |
+
```
|
| 176 |
+
|
| 177 |
+
### 2. Synthesis Report (์ต์ข
๋ฆฌํฌํธ)
|
| 178 |
|
| 179 |
```markdown
|
| 180 |
+
# EOD ๋ฆฌํฌํธ: ์ผ์ฑ์ ์ (005930.KS)
|
| 181 |
+
|
| 182 |
## Executive Summary
|
| 183 |
[2-3๋ฌธ์ฅ ํต์ฌ ์์ฝ]
|
| 184 |
|
| 185 |
## Key Metrics
|
| 186 |
| ์งํ | ์์น |
|
| 187 |
|------|------|
|
|
|
|
|
|
|
|
|
|
| 188 |
|
| 189 |
## ์์ฅ ๋ํฅ ๋ถ์
|
| 190 |
+
## ์๊ธ ๋ถ์
|
| 191 |
+
## ๋ด์ค ์ํฅ ๋ถ์
|
| 192 |
## ๋ฆฌ์คํฌ ์์ธ
|
|
|
|
|
|
|
|
|
|
| 193 |
## ํฌ์ ์๊ฒฌ
|
|
|
|
| 194 |
```
|
| 195 |
|
| 196 |
## Training Data
|
|
|
|
| 202 |
| Reasoning Traces | 5K | ์ฌ๊ณ ๊ณผ์ ํ์ต |
|
| 203 |
| DPO Pairs | 7.7K | ์ ํธ๋ ์ ๋ ฌ |
|
| 204 |
|
| 205 |
+
## DPO Improvements
|
|
|
|
|
|
|
| 206 |
|
| 207 |
+
- โ
**์ค๊ตญ์ด leak ์ ๊ฑฐ**: Stress test 10/10 CLEAN
|
| 208 |
- โ
**์์ด leak ๊ฐ์**: ๋ถํ์ํ ์์ด ์ฌ์ฉ ์ต์ํ
|
| 209 |
+
- โ
**ํ์ ์ค์**: Reasoning Trace JSON + 7-section Report
|
| 210 |
- โ
**ํ๊ตญ์ด ํ์ง**: ์์ฐ์ค๋ฌ์ด ํ๊ตญ์ด ํํ
|
| 211 |
|
| 212 |
## Limitations
|
|
|
|
| 214 |
- ์ค์๊ฐ ์์ธ ๋ฐ์ดํฐ ์ ๊ทผ ๋ถ๊ฐ (์ธ๋ถ API ํ์)
|
| 215 |
- ํฌ์ ์กฐ์ธ์ด ์๋ ์ ๋ณด ์ ๊ณต ๋ชฉ์
|
| 216 |
- 8K ์ปจํ
์คํธ ์ ํ์ผ๋ก ๊ธด ๋ฌธ์ ์ฒ๋ฆฌ ํ๊ณ
|
| 217 |
+
- ํ ๋ฃจ์๋ค์ด์
์์น ๊ฐ๋ฅ (์์น ๋ฐ์ดํฐ๋ ์ธ๋ถ ๊ฒ์ฆ ํ์)
|
| 218 |
|
| 219 |
## Citation
|
| 220 |
|
|
|
|
| 224 |
author={intrect},
|
| 225 |
year={2026},
|
| 226 |
publisher={Hugging Face},
|
| 227 |
+
url={https://huggingface.co/intrect/VELA}
|
| 228 |
}
|
| 229 |
```
|
| 230 |
|
|
|
|
| 232 |
|
| 233 |
| ๋ฒ์ | ๋ ์ง | ๋ณ๊ฒฝ์ฌํญ |
|
| 234 |
|------|------|----------|
|
| 235 |
+
| v1.1 | 2026-02-12 | GGUF ์์ํ ๋ชจ๋ธ ์ถ๊ฐ (Q4_K_M, Q8_0), ๋ฒค์น๋งํฌ |
|
| 236 |
+
| v1.0 | 2026-01-28 | DPO ๋ณํฉ, ์ค๊ตญ์ด/์์ด leak ํด๊ฒฐ |
|
| 237 |
+
| v0.9 | 2026-01-15 | SFT ๋ฒ ์ด์ค ๋ชจ๋ธ ๊ณต๊ฐ |
|
| 238 |
|
| 239 |
---
|
| 240 |
|