File size: 10,667 Bytes
7b7257a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 | """KAIdol 학습 모델 레지스트리 - 모든 학습된 모델 정의"""
from typing import Dict, List, Optional
# 기본 모델 정보 (HuggingFace Hub)
BASE_MODELS = {
"hyperclovax-32b": "naver-hyperclovax/HyperCLOVAX-SEED-Think-32B",
"qwen2.5-72b": "Qwen/Qwen2.5-72B-Instruct",
"qwen2.5-32b": "Qwen/Qwen2.5-32B-Instruct",
"qwen2.5-14b": "Qwen/Qwen2.5-14B-Instruct",
"qwen2.5-7b": "Qwen/Qwen2.5-7B-Instruct",
"qwen3-8b": "Qwen/Qwen3-8B",
"exaone-7.8b": "LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct",
"solar-10.7b": "upstage/SOLAR-10.7B-Instruct-v1.0",
"solar-pro": "upstage/solar-pro-preview-instruct",
"varco-8b": "NCSOFT/Llama-VARCO-8B-Instruct",
"kanana-2-30b-thinking": "kakaocorp/kanana-2-30b-a3b-thinking",
"kanana-2-30b-instruct": "kakaocorp/kanana-2-30b-a3b-instruct",
"llama-3.3-70b": "meta-llama/Llama-3.3-70B-Instruct",
}
# 전체 모델 레지스트리
MODEL_REGISTRY: Dict[str, Dict[str, Dict]] = {
# ============================================================
# DPO v5 계열 (2026-01-13)
# ============================================================
"dpo-v5": {
"hyperclovax-32b-dpo-v5": {
"path": "outputs/dpo_v5/hyperclovax-32b-dpo-v5-20260113-0012",
"base": BASE_MODELS["hyperclovax-32b"],
"method": "DPO",
"size": "32B",
"description": "HyperCLOVAX 32B DPO v5 (Primary)",
"recommended": True,
},
"qwen2.5-14b-dpo-v5": {
"path": "outputs/dpo_v5/qwen2.5-14b-dpo-v5-20260113-0045",
"base": BASE_MODELS["qwen2.5-14b"],
"method": "DPO",
"size": "14B",
"description": "Qwen2.5 14B DPO v5",
},
"qwen2.5-7b-dpo-v5": {
"path": "outputs/dpo_v5/qwen2.5-7b-dpo-v5-20260113-0052",
"base": BASE_MODELS["qwen2.5-7b"],
"method": "DPO",
"size": "7B",
"description": "Qwen2.5 7B DPO v5",
},
"exaone-7.8b-dpo-v5": {
"path": "outputs/dpo_v5/exaone-7.8b-dpo-v5-20260113-0052",
"base": BASE_MODELS["exaone-7.8b"],
"method": "DPO",
"size": "7.8B",
"description": "EXAONE 7.8B DPO v5",
},
"qwen3-8b-dpo-v5": {
"path": "outputs/dpo_v5/qwen3-8b-dpo-v5-20260113-0052",
"base": BASE_MODELS["qwen3-8b"],
"method": "DPO",
"size": "8B",
"description": "Qwen3 8B DPO v5",
},
"solar-10.7b-dpo-v5": {
"path": "outputs/dpo_v5/solar-10.7b-dpo-v5-20260113-0045",
"base": BASE_MODELS["solar-10.7b"],
"method": "DPO",
"size": "10.7B",
"description": "Solar 10.7B DPO v5",
},
},
# ============================================================
# SFT Thinking 계열 (2026-01-16)
# ============================================================
"sft-thinking": {
"qwen2.5-14b-thinking": {
"path": "outputs/qwen2.5-14b-thinking-full",
"base": BASE_MODELS["qwen2.5-14b"],
"method": "SFT",
"size": "14B",
"description": "Qwen2.5 14B SFT Thinking",
},
"qwen2.5-7b-thinking": {
"path": "outputs/qwen2.5-7b-thinking-full",
"base": BASE_MODELS["qwen2.5-7b"],
"method": "SFT",
"size": "7B",
"description": "Qwen2.5 7B SFT Thinking",
},
"exaone-7.8b-thinking": {
"path": "outputs/exaone-7.8b-thinking-full",
"base": BASE_MODELS["exaone-7.8b"],
"method": "SFT",
"size": "7.8B",
"description": "EXAONE 7.8B SFT Thinking",
},
},
# ============================================================
# Phase 7 Students (Kimi K2 Distillation)
# ============================================================
"phase7-students": {
"kanana-30b-thinking-kimi": {
"path": "outputs/phase7_students/kanana-2-30b-thinking-kimi-student",
"base": BASE_MODELS["kanana-2-30b-thinking"],
"method": "Distillation",
"size": "30B (3B active)",
"description": "Kanana 30B Thinking Kimi Student",
},
"kanana-30b-instruct-kimi": {
"path": "outputs/phase7_students/kanana-2-30b-instruct-kimi-student",
"base": BASE_MODELS["kanana-2-30b-instruct"],
"method": "Distillation",
"size": "30B (3B active)",
"description": "Kanana 30B Instruct Kimi Student",
},
"qwen2.5-14b-kimi": {
"path": "outputs/phase7_students/qwen2.5-14b-kimi-student",
"base": BASE_MODELS["qwen2.5-14b"],
"method": "Distillation",
"size": "14B",
"description": "Qwen2.5 14B Kimi Student",
},
"qwen2.5-7b-kimi-v3": {
"path": "outputs/phase7_students/qwen2.5-7b-kimi-student-v3",
"base": BASE_MODELS["qwen2.5-7b"],
"method": "Distillation",
"size": "7B",
"description": "Qwen2.5 7B Kimi Student v3",
},
"exaone-7.8b-kimi": {
"path": "outputs/phase7_students/exaone-7.8b-kimi-student",
"base": BASE_MODELS["exaone-7.8b"],
"method": "Distillation",
"size": "7.8B",
"description": "EXAONE 7.8B Kimi Student",
},
},
# ============================================================
# V7 Students (Latest - 2026-01-17~19)
# ============================================================
"v7-students": {
"qwen2.5-72b-v7": {
"path": "outputs/v7_students/qwen2.5-72b-v7-20260119-1113",
"base": BASE_MODELS["qwen2.5-72b"],
"method": "SFT",
"size": "72B",
"description": "Qwen2.5 72B V7 (Latest)",
},
"llama-3.3-70b-v7": {
"path": "outputs/v7_students/llama-3.3-70b-v7-20260119-1114",
"base": BASE_MODELS["llama-3.3-70b"],
"method": "SFT",
"size": "70B",
"description": "Llama 3.3 70B V7 (Latest)",
},
"qwen2.5-32b-v7": {
"path": "outputs/v7_students/qwen2.5-32b-v7-20260118-1135",
"base": BASE_MODELS["qwen2.5-32b"],
"method": "SFT",
"size": "32B",
"description": "Qwen2.5 32B V7",
},
"qwen2.5-14b-v7": {
"path": "outputs/v7_students/qwen2.5-14b-v7-20260118-1135",
"base": BASE_MODELS["qwen2.5-14b"],
"method": "SFT",
"size": "14B",
"description": "Qwen2.5 14B V7",
},
"qwen2.5-7b-v7": {
"path": "outputs/v7_students/qwen2.5-7b-v7-20260118-1135",
"base": BASE_MODELS["qwen2.5-7b"],
"method": "SFT",
"size": "7B",
"description": "Qwen2.5 7B V7",
},
"exaone-7.8b-v7": {
"path": "outputs/v7_students/exaone-7.8b-v7-20260118-1135",
"base": BASE_MODELS["exaone-7.8b"],
"method": "SFT",
"size": "7.8B",
"description": "EXAONE 7.8B V7",
},
"qwen3-8b-v7": {
"path": "outputs/v7_students/qwen3-8b-v7-20260118-1135",
"base": BASE_MODELS["qwen3-8b"],
"method": "SFT",
"size": "8B",
"description": "Qwen3 8B V7",
},
"solar-pro-v7": {
"path": "outputs/v7_students/solar-pro-v7-20260118-1135",
"base": BASE_MODELS["solar-pro"],
"method": "SFT",
"size": "22B",
"description": "Solar Pro V7",
},
"varco-8b-v7": {
"path": "outputs/v7_students/varco-8b-v7-20260118-1135",
"base": BASE_MODELS["varco-8b"],
"method": "SFT",
"size": "8B",
"description": "VARCO 8B V7",
},
},
# ============================================================
# 기타 학습 모델 (DPO, etc.)
# ============================================================
"others": {
"exaone-7.8b-dpo": {
"path": "outputs/exaone-7.8b-dpo",
"base": BASE_MODELS["exaone-7.8b"],
"method": "DPO",
"size": "7.8B",
"description": "EXAONE 7.8B DPO (Standalone)",
},
"qwen2.5-7b-dpo": {
"path": "outputs/qwen2.5-7b-dpo",
"base": BASE_MODELS["qwen2.5-7b"],
"method": "DPO",
"size": "7B",
"description": "Qwen2.5 7B DPO (Standalone)",
},
},
}
def get_all_models() -> List[str]:
"""모든 모델 ID 목록 반환"""
models = []
for category, model_dict in MODEL_REGISTRY.items():
models.extend(model_dict.keys())
return models
def get_model_info(model_id: str) -> Optional[Dict]:
"""모델 ID로 정보 조회"""
for category, model_dict in MODEL_REGISTRY.items():
if model_id in model_dict:
info = model_dict[model_id].copy()
info["category"] = category
info["id"] = model_id
return info
return None
def get_models_by_category(category: str) -> List[str]:
"""카테고리별 모델 목록"""
return list(MODEL_REGISTRY.get(category, {}).keys())
def get_all_categories() -> List[str]:
"""모든 카테고리 목록"""
return list(MODEL_REGISTRY.keys())
def get_models_for_dropdown() -> List[tuple]:
"""드롭다운용 (display_name, model_id) 튜플 리스트"""
result = []
for category, model_dict in MODEL_REGISTRY.items():
for model_id, info in model_dict.items():
display = f"[{info.get('size', '?')}] {info.get('description', model_id)}"
result.append((display, model_id))
return result
def get_small_models(max_size_gb: int = 16) -> List[str]:
"""메모리 제한에 맞는 소형 모델만 반환 (4bit 양자화 기준)"""
# 4bit 양자화 시 대략적인 메모리: 7B~2GB, 14B~4GB, 32B~8GB, 72B~18GB
size_map = {
"7B": 2, "7.8B": 2, "8B": 2,
"10.7B": 3, "14B": 4, "22B": 6,
"30B (3B active)": 1, # MoE
"32B": 8, "70B": 18, "72B": 18,
}
result = []
for model_id in get_all_models():
info = get_model_info(model_id)
if info:
size_str = info.get("size", "72B")
estimated_gb = size_map.get(size_str, 20)
if estimated_gb <= max_size_gb:
result.append(model_id)
return result
|