Spaces:
Running
Running
File size: 10,272 Bytes
173f28e d6ca6d1 173f28e f56dbf3 7c4c461 173f28e 7c4c461 173f28e 7c4c461 173f28e 7c4c461 173f28e bf74331 173f28e d6ca6d1 | 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 | from __future__ import annotations
import json
from dataclasses import asdict, dataclass
from pathlib import Path
@dataclass
class ModelConfig:
name: str
model_id: str
is_baseline: bool = False
backend: str = "sbert"
gguf_file: str | None = None
# Prefixes prepended to inputs at encode time. Many retrieval-tuned
# encoders (E5, BGE, Nomic, mxbai, Arctic, ...) expect distinct
# query/passage instructions for best quality.
query_prefix: str = ""
passage_prefix: str = ""
# Some models (e.g. Harrier) ship named prompts in the sentence-transformers
# config and should be invoked via encode(..., prompt_name=...) rather than
# by manually prepending a string. Only used by the sbert backend.
query_prompt_name: str | None = None
passage_prompt_name: str | None = None
# Some HF repos ship custom modeling code and require opting in.
trust_remote_code: bool = False
REGISTRY: dict[str, ModelConfig] = {
# --- Official sentence-transformers models ---
"mpnet": ModelConfig(
name="all-mpnet-base-v2",
model_id="sentence-transformers/all-mpnet-base-v2",
is_baseline=True,
),
"minilm-l6": ModelConfig(
name="all-MiniLM-L6-v2",
model_id="sentence-transformers/all-MiniLM-L6-v2",
),
"minilm-l12": ModelConfig(
name="all-MiniLM-L12-v2",
model_id="sentence-transformers/all-MiniLM-L12-v2",
),
"distilroberta": ModelConfig(
name="all-distilroberta-v1",
model_id="sentence-transformers/all-distilroberta-v1",
),
"multi-qa-mpnet": ModelConfig(
name="multi-qa-mpnet-base-dot-v1",
model_id="sentence-transformers/multi-qa-mpnet-base-dot-v1",
),
"multi-qa-distilbert": ModelConfig(
name="multi-qa-distilbert-cos-v1",
model_id="sentence-transformers/multi-qa-distilbert-cos-v1",
),
"multi-qa-minilm": ModelConfig(
name="multi-qa-MiniLM-L6-cos-v1",
model_id="sentence-transformers/multi-qa-MiniLM-L6-cos-v1",
),
"paraphrase-mpnet": ModelConfig(
name="paraphrase-mpnet-base-v2",
model_id="sentence-transformers/paraphrase-mpnet-base-v2",
),
"paraphrase-minilm-l6": ModelConfig(
name="paraphrase-MiniLM-L6-v2",
model_id="sentence-transformers/paraphrase-MiniLM-L6-v2",
),
"paraphrase-minilm-l3": ModelConfig(
name="paraphrase-MiniLM-L3-v2",
model_id="sentence-transformers/paraphrase-MiniLM-L3-v2",
),
"paraphrase-multilingual-mpnet": ModelConfig(
name="paraphrase-multilingual-mpnet-base-v2",
model_id="sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
),
"paraphrase-multilingual-minilm": ModelConfig(
name="paraphrase-multilingual-MiniLM-L12-v2",
model_id="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
),
"distiluse-multilingual": ModelConfig(
name="distiluse-base-multilingual-cased-v2",
model_id="sentence-transformers/distiluse-base-multilingual-cased-v2",
),
"labse": ModelConfig(
name="LaBSE",
model_id="sentence-transformers/LaBSE",
),
# --- BGE (BAAI) ---
# v1.5 English BGE models: only the query is instructed; passages are raw.
"bge-small": ModelConfig(
name="bge-small-en-v1.5",
model_id="BAAI/bge-small-en-v1.5",
query_prefix="Represent this sentence for searching relevant passages: ",
),
"bge-base": ModelConfig(
name="bge-base-en-v1.5",
model_id="BAAI/bge-base-en-v1.5",
query_prefix="Represent this sentence for searching relevant passages: ",
),
"bge-large": ModelConfig(
name="bge-large-en-v1.5",
model_id="BAAI/bge-large-en-v1.5",
query_prefix="Represent this sentence for searching relevant passages: ",
),
# bge-m3 does not use an instruction prefix.
"bge-m3": ModelConfig(
name="bge-m3",
model_id="BAAI/bge-m3",
),
# --- E5 (intfloat) ---
# All E5 models expect "query: " / "passage: " prefixes.
"e5-small-v2": ModelConfig(
name="e5-small-v2",
model_id="intfloat/e5-small-v2",
query_prefix="query: ",
passage_prefix="passage: ",
),
"e5-base-v2": ModelConfig(
name="e5-base-v2",
model_id="intfloat/e5-base-v2",
query_prefix="query: ",
passage_prefix="passage: ",
),
"e5-large-v2": ModelConfig(
name="e5-large-v2",
model_id="intfloat/e5-large-v2",
query_prefix="query: ",
passage_prefix="passage: ",
),
"multilingual-e5-small": ModelConfig(
name="multilingual-e5-small",
model_id="intfloat/multilingual-e5-small",
query_prefix="query: ",
passage_prefix="passage: ",
),
"multilingual-e5-base": ModelConfig(
name="multilingual-e5-base",
model_id="intfloat/multilingual-e5-base",
query_prefix="query: ",
passage_prefix="passage: ",
),
"multilingual-e5-large": ModelConfig(
name="multilingual-e5-large",
model_id="intfloat/multilingual-e5-large",
query_prefix="query: ",
passage_prefix="passage: ",
),
# --- GTE (Alibaba) ---
# GTE v1 models do not use prefixes.
"gte-small": ModelConfig(
name="gte-small",
model_id="thenlper/gte-small",
),
"gte-base": ModelConfig(
name="gte-base",
model_id="thenlper/gte-base",
),
"gte-large": ModelConfig(
name="gte-large",
model_id="thenlper/gte-large",
),
# --- Nomic ---
# Nomic expects task-specific prefixes.
"nomic-v1": ModelConfig(
name="nomic-embed-text-v1",
model_id="nomic-ai/nomic-embed-text-v1",
query_prefix="search_query: ",
passage_prefix="search_document: ",
trust_remote_code=True,
),
"nomic-v1.5": ModelConfig(
name="nomic-embed-text-v1.5",
model_id="nomic-ai/nomic-embed-text-v1.5",
query_prefix="search_query: ",
passage_prefix="search_document: ",
trust_remote_code=True,
),
# --- Mixedbread ---
# mxbai instructs the query only.
"mxbai-large": ModelConfig(
name="mxbai-embed-large-v1",
model_id="mixedbread-ai/mxbai-embed-large-v1",
query_prefix="Represent this sentence for searching relevant passages: ",
),
# --- Jina ---
# Jina v2 does not use prefixes.
"jina-v2-small": ModelConfig(
name="jina-embeddings-v2-small-en",
model_id="jinaai/jina-embeddings-v2-small-en",
trust_remote_code=True,
),
"jina-v2-base": ModelConfig(
name="jina-embeddings-v2-base-en",
model_id="jinaai/jina-embeddings-v2-base-en",
trust_remote_code=True,
),
# --- Microsoft Harrier-OSS-v1 ---
# Decoder-only multilingual embedding family. Queries use the
# `web_search_query` named prompt (shipped in the model's ST config);
# passages are encoded raw.
"harrier-270m": ModelConfig(
name="harrier-oss-v1-270m",
model_id="microsoft/harrier-oss-v1-270m",
query_prompt_name="web_search_query",
),
"harrier-0.6b": ModelConfig(
name="harrier-oss-v1-0.6b",
model_id="microsoft/harrier-oss-v1-0.6b",
query_prompt_name="web_search_query",
),
"harrier-27b": ModelConfig(
name="harrier-oss-v1-27b",
model_id="microsoft/harrier-oss-v1-27b",
query_prompt_name="web_search_query",
),
# --- Snowflake Arctic ---
# Arctic instructs the query only.
"arctic-xs": ModelConfig(
name="snowflake-arctic-embed-xs",
model_id="Snowflake/snowflake-arctic-embed-xs",
query_prefix="Represent this sentence for searching relevant passages: ",
),
"arctic-s": ModelConfig(
name="snowflake-arctic-embed-s",
model_id="Snowflake/snowflake-arctic-embed-s",
query_prefix="Represent this sentence for searching relevant passages: ",
),
"arctic-m": ModelConfig(
name="snowflake-arctic-embed-m",
model_id="Snowflake/snowflake-arctic-embed-m",
query_prefix="Represent this sentence for searching relevant passages: ",
),
"arctic-l": ModelConfig(
name="snowflake-arctic-embed-l",
model_id="Snowflake/snowflake-arctic-embed-l",
query_prefix="Represent this sentence for searching relevant passages: ",
),
# "bge-small-fe": ModelConfig(
# name="bge-small-en-v1.5 (fastembed)",
# model_id="BAAI/bge-small-en-v1.5",
# backend="fastembed",
# ),
# "all-minilm-fe": ModelConfig(
# name="all-MiniLM-L6-v2 (fastembed)",
# model_id="sentence-transformers/all-MiniLM-L6-v2",
# backend="fastembed",
# ),
# "bge-small-le": ModelConfig(
# name="bge-small-en-v1.5 (libembedding)",
# model_id="BAAI/bge-small-en-v1.5",
# backend="libembedding",
# ),
# "all-minilm-le": ModelConfig(
# name="all-MiniLM-L6-v2 (libembedding)",
# model_id="sentence-transformers/all-MiniLM-L6-v2",
# backend="libembedding",
# ),
}
VALID_BACKENDS = {"sbert", "fastembed", "libembedding", "gguf"}
CUSTOM_MODELS_PATH = Path(__file__).parent / "custom_models.json"
def register_model(key: str, config: ModelConfig) -> None:
if key in REGISTRY:
raise ValueError(f"Model key '{key}' already exists in registry")
if config.backend not in VALID_BACKENDS:
raise ValueError(f"Invalid backend '{config.backend}'. Must be one of: {VALID_BACKENDS}")
REGISTRY[key] = config
def load_custom_models_from_file(path: Path = CUSTOM_MODELS_PATH) -> None:
if not path.exists():
return
with open(path) as f:
entries = json.load(f)
for key, fields in entries.items():
if key not in REGISTRY:
REGISTRY[key] = ModelConfig(**fields)
def save_custom_model_to_file(key: str, config: ModelConfig, path: Path = CUSTOM_MODELS_PATH) -> None:
existing: dict = {}
if path.exists():
with open(path) as f:
existing = json.load(f)
existing[key] = asdict(config)
with open(path, "w") as f:
json.dump(existing, f, indent=2)
|