README.md · bge-reranker-base

1

---

2

license: mit

3

language:

4

- en

5

- zh

6

tags:

7

- mteb

8

- text-embeddings-inference

9

model-index:

10

- name: bge-reranker-base

11

results:

12

- task:

13

type: Reranking

14

dataset:

15

type: C-MTEB/CMedQAv1-reranking

16

name: MTEB CMedQAv1

17

config: default

18

split: test

19

revision: None

20

metrics:

21

- type: map

22

value: 81.27206722525007

23

- type: mrr

24

value: 84.14238095238095

25

- task:

26

type: Reranking

27

dataset:

28

type: C-MTEB/CMedQAv2-reranking

29

name: MTEB CMedQAv2

30

config: default

31

split: test

32

revision: None

33

metrics:

34

- type: map

35

value: 84.10369934291236

36

- type: mrr

37

value: 86.79376984126984

38

- task:

39

type: Reranking

40

dataset:

41

type: C-MTEB/Mmarco-reranking

42

name: MTEB MMarcoReranking

43

config: default

44

split: dev

45

revision: None

46

metrics:

47

- type: map

48

value: 35.4600511272538

49

- type: mrr

50

value: 34.60238095238095

51

- task:

52

type: Reranking

53

dataset:

54

type: C-MTEB/T2Reranking

55

name: MTEB T2Reranking

56

config: default

57

split: dev

58

revision: None

59

metrics:

60

- type: map

61

value: 67.27728847727172

62

- type: mrr

63

value: 77.1315192743764

64

pipeline_tag: text-classification

65

library_name: sentence-transformers

66

---

67

68

**We have updated the [new reranker](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_reranker), supporting larger lengths, more languages, and achieving better performance.**

69

70

<h1 align="center">FlagEmbedding</h1>

71

72

73

<h4 align="center">

74

<p>

75

<a href=#model-list>Model List</a> |

76

<a href=#frequently-asked-questions>FAQ</a> |

77

<a href=#usage>Usage</a> |

78

<a href="#evaluation">Evaluation</a> |

79

<a href="#train">Train</a> |

80

<a href="#citation">Citation</a> |

81

<a href="#license">License</a>

82

<p>

83

</h4>

84

85

**More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).**

86

87

88

[English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)

89

90

91

FlagEmbedding focuses on retrieval-augmented LLMs, consisting of the following projects currently:

92

93

- **Long-Context LLM**: [Activation Beacon](https://github.com/FlagOpen/FlagEmbedding/tree/master/Long_LLM/activation_beacon)

94

- **Fine-tuning of LM** : [LM-Cocktail](https://github.com/FlagOpen/FlagEmbedding/tree/master/LM_Cocktail)

95

- **Embedding Model**: [Visualized-BGE](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/visual), [BGE-M3](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3), [LLM Embedder](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_embedder), [BGE Embedding](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/baai_general_embedding)

96

- **Reranker Model**: [llm rerankers](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_reranker), [BGE Reranker](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/reranker)

97

- **Benchmark**: [C-MTEB](https://github.com/FlagOpen/FlagEmbedding/tree/master/C_MTEB)

98

99

## News

100

- 3/18/2024: Release new [rerankers](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_reranker), built upon powerful M3 and LLM (GEMMA and MiniCPM, not so large actually) backbones, supporitng multi-lingual processing and larger inputs, massive improvements of ranking performances on BEIR, C-MTEB/Retrieval, MIRACL, LlamaIndex Evaluation.

101

- 3/18/2024: Release [Visualized-BGE](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/visual), equipping BGE with visual capabilities. Visualized-BGE can be utilized to generate embeddings for hybrid image-text data.

102

- 1/30/2024: Release **BGE-M3**, a new member to BGE model series! M3 stands for **M**ulti-linguality (100+ languages), **M**ulti-granularities (input length up to 8192), **M**ulti-Functionality (unification of dense, lexical, multi-vec/colbert retrieval).

103

It is the first embedding model which supports all three retrieval methods, achieving new SOTA on multi-lingual (MIRACL) and cross-lingual (MKQA) benchmarks.

104

[Technical Report](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/BGE_M3/BGE_M3.pdf) and [Code](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3). :fire:

105

- 1/9/2024: Release [Activation-Beacon](https://github.com/FlagOpen/FlagEmbedding/tree/master/Long_LLM/activation_beacon), an effective, efficient, compatible, and low-cost (training) method to extend the context length of LLM. [Technical Report](https://arxiv.org/abs/2401.03462) :fire:

106

- 12/24/2023: Release **LLaRA**, a LLaMA-7B based dense retriever, leading to state-of-the-art performances on MS MARCO and BEIR. Model and code will be open-sourced. Please stay tuned. [Technical Report](https://arxiv.org/abs/2312.15503)

107

- 11/23/2023: Release [LM-Cocktail](https://github.com/FlagOpen/FlagEmbedding/tree/master/LM_Cocktail), a method to maintain general capabilities during fine-tuning by merging multiple language models. [Technical Report](https://arxiv.org/abs/2311.13534) :fire:

108

- 10/12/2023: Release [LLM-Embedder](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_embedder), a unified embedding model to support diverse retrieval augmentation needs for LLMs. [Technical Report](https://arxiv.org/pdf/2310.07554.pdf)

109

- 09/15/2023: The [technical report](https://arxiv.org/pdf/2309.07597.pdf) of BGE has been released

110

- 09/15/2023: The [massive training data](https://data.baai.ac.cn/details/BAAI-MTP) of BGE has been released

111

- 09/12/2023: New models:

112

    - **New reranker model**: release cross-encoder models `BAAI/bge-reranker-base` and `BAAI/bge-reranker-large`, which are more powerful than embedding model. We recommend to use/fine-tune them to re-rank top-k documents returned by embedding models.

113

    - **update embedding model**: release `bge-*-v1.5` embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.

114

115

116

<details>

117

<summary>More</summary>

118

119

120

- 09/07/2023: Update [fine-tune code](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md): Add script to mine hard negatives and support adding instruction during fine-tuning.

121

- 08/09/2023: BGE Models are integrated into **Langchain**, you can use it like [this](#using-langchain); C-MTEB **leaderboard** is [available](https://huggingface.co/spaces/mteb/leaderboard).

122

- 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**

123

- 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:

124

- 08/01/2023: We release the [Chinese Massive Text Embedding Benchmark](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB) (**C-MTEB**), consisting of 31 test dataset.

125

126

</details>

127

128

129

## Model List

130

131

`bge` is short for `BAAI general embedding`.

132

133

134

|:-------------------------------|:--------:| :--------:| :--------:|:--------:|

135

| [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3)                   |    Multilingual     |    [Inference](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3#usage) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3)    | Multi-Functionality(dense retrieval, sparse retrieval, multi-vector(colbert)), Multi-Linguality, and Multi-Granularity(8192 tokens) |  |

136

|  [BAAI/llm-embedder](https://huggingface.co/BAAI/llm-embedder)  |   English | [Inference](./FlagEmbedding/llm_embedder/README.md) [Fine-tune](./FlagEmbedding/llm_embedder/README.md) | a unified embedding model to support diverse retrieval augmentation needs for LLMs | See [README](./FlagEmbedding/llm_embedder/README.md) |

137

|  [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large)  |   Chinese and English | [Inference](#usage-for-reranker) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/reranker) | a cross-encoder model which is more accurate but less efficient [2] |   |

138

|  [BAAI/bge-reranker-base](https://huggingface.co/BAAI/bge-reranker-base) |   Chinese and English | [Inference](#usage-for-reranker) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/reranker) | a cross-encoder model which is more accurate but less efficient [2] |   |

139

|  [BAAI/bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution | `Represent this sentence for searching relevant passages: `  |

140

|  [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution | `Represent this sentence for searching relevant passages: `  |

141

|  [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution  | `Represent this sentence for searching relevant passages: `  |

142

|  [BAAI/bge-large-zh-v1.5](https://huggingface.co/BAAI/bge-large-zh-v1.5) |   Chinese | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution | `为这个句子生成表示以用于检索相关文章：`  |

143

|  [BAAI/bge-base-zh-v1.5](https://huggingface.co/BAAI/bge-base-zh-v1.5) |   Chinese |  [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution | `为这个句子生成表示以用于检索相关文章：`  |

144

|  [BAAI/bge-small-zh-v1.5](https://huggingface.co/BAAI/bge-small-zh-v1.5) |   Chinese | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | version 1.5 with more reasonable similarity distribution | `为这个句子生成表示以用于检索相关文章：`  |

145

|  [BAAI/bge-large-en](https://huggingface.co/BAAI/bge-large-en) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | :trophy: rank **1st** in [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard | `Represent this sentence for searching relevant passages: `  |

146

|  [BAAI/bge-base-en](https://huggingface.co/BAAI/bge-base-en) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | a base-scale model but with similar ability to `bge-large-en` | `Represent this sentence for searching relevant passages: `  |

147

|  [BAAI/bge-small-en](https://huggingface.co/BAAI/bge-small-en) |   English | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) |a small-scale model but with competitive performance  | `Represent this sentence for searching relevant passages: `  |

148

|  [BAAI/bge-large-zh](https://huggingface.co/BAAI/bge-large-zh) |   Chinese | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | :trophy: rank **1st** in [C-MTEB](https://github.com/FlagOpen/FlagEmbedding/tree/master/C_MTEB) benchmark | `为这个句子生成表示以用于检索相关文章：`  |

149

|  [BAAI/bge-base-zh](https://huggingface.co/BAAI/bge-base-zh) |   Chinese |  [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | a base-scale model but with similar ability to `bge-large-zh` | `为这个句子生成表示以用于检索相关文章：`  |

150

|  [BAAI/bge-small-zh](https://huggingface.co/BAAI/bge-small-zh) |   Chinese | [Inference](#usage-for-embedding-model) [Fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) | a small-scale model but with competitive performance | `为这个句子生成表示以用于检索相关文章：`  |

151

152

153

[1\]: If you need to search the relevant passages to a query, we suggest to add the instruction to the query; in other cases, no instruction is needed, just use the original query directly. In all cases, **no instruction** needs to be added to passages.

154

155

[2\]: Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding. To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models.

156

For examples, use bge embedding model to retrieve top 100 relevant documents, and then use bge reranker to re-rank the top 100 document to get the final top-3 results.

157

158

All models have been uploaded to Huggingface Hub, and you can see them at https://huggingface.co/BAAI.

159

If you cannot open the Huggingface Hub, you also can download the models at https://model.baai.ac.cn/models .

160

161

162

## Frequently asked questions

163

164

<details>

165

<summary>1. How to fine-tune bge embedding model?</summary>

166

167

168

Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.

169

Some suggestions:

170

- Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#hard-negatives), which can improve the retrieval performance.

171

- If you pre-train bge on your data, the pre-trained model cannot be directly used to calculate similarity, and it must be fine-tuned with contrastive learning before computing similarity.

172

- If the accuracy of the fine-tuned model is still not high, it is recommended to use/fine-tune the cross-encoder model (bge-reranker) to re-rank top-k results.

173

  Hard negatives also are needed to fine-tune reranker. Refer to this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/reranker) for the fine-tuning for reranker

174

175

176

</details>

177

178

<details>

179

<summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>

180

181

182

**Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**

183

184

Since we finetune the models by contrastive learning with a temperature of 0.01,

185

the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].

186

So a similarity score greater than 0.5 does not indicate that the two sentences are similar.

187

188

For downstream tasks, such as passage retrieval or semantic similarity,

189

**what matters is the relative order of the scores, not the absolute value.**

190

If you need to filter similar sentences based on a similarity threshold,

191

please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).

192

193

</details>

194

195

<details>

196

<summary>3. When does the query instruction need to be used</summary>

197

198

199

200

For the `bge-*-v1.5`, we improve its retrieval ability when not using instruction.

201

No instruction only has a slight degradation in retrieval performance compared with using instruction.

202

So you can generate embedding without instruction in all cases for convenience.

203

204

For a retrieval task that uses short queries to find long related documents,

205

it is recommended to add instructions for these short queries.

206

**The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**

207

In all cases, the documents/passages do not need to add the instruction.

208

209

</details>

210

211

212

## Usage

213

214

### Usage for Embedding Model

215

216

Here are some examples for using `bge` models with

217

[FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).

218

219

#### Using FlagEmbedding

220

```

221

pip install -U FlagEmbedding

222

```

223

If it doesn't work for you, you can see [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md) for more methods to install FlagEmbedding.

224

225

```python

226

from FlagEmbedding import FlagModel

227

sentences_1 = ["样例数据-1", "样例数据-2"]

228

sentences_2 = ["样例数据-3", "样例数据-4"]

229

model = FlagModel('BAAI/bge-large-zh-v1.5',

230

query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章：",

231

use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation

232

embeddings_1 = model.encode(sentences_1)

233

embeddings_2 = model.encode(sentences_2)

234

similarity = embeddings_1 @ embeddings_2.T

235

print(similarity)

236

237

# for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query

238

# corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction

239

queries = ['query_1', 'query_2']

240

passages = ["样例文档-1", "样例文档-2"]

241

q_embeddings = model.encode_queries(queries)

242

p_embeddings = model.encode(passages)

243

scores = q_embeddings @ p_embeddings.T

244

```

245

For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).

246

247

By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.

248

You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.

249

250

251

#### Using Sentence-Transformers

252

253

You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):

254

255

```

256

pip install -U sentence-transformers

257

```

258

```python

259

from sentence_transformers import SentenceTransformer

260

sentences_1 = ["样例数据-1", "样例数据-2"]

261

sentences_2 = ["样例数据-3", "样例数据-4"]

262

model = SentenceTransformer('BAAI/bge-large-zh-v1.5')

263

embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)

264

embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)

265

similarity = embeddings_1 @ embeddings_2.T

266

print(similarity)

267

```

268

For s2p(short query to long passage) retrieval task,

269

each short query should start with an instruction (instructions see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list)).

270

But the instruction is not needed for passages.

271

```python

272

from sentence_transformers import SentenceTransformer

273

queries = ['query_1', 'query_2']

274

passages = ["样例文档-1", "样例文档-2"]

275

instruction = "为这个句子生成表示以用于检索相关文章："

276

277

model = SentenceTransformer('BAAI/bge-large-zh-v1.5')

278

q_embeddings = model.encode([instruction+q for q in queries], normalize_embeddings=True)

279

p_embeddings = model.encode(passages, normalize_embeddings=True)

280

scores = q_embeddings @ p_embeddings.T

281

```

282

283

#### Using Langchain

284

285

You can use `bge` in langchain like this:

286

```python

287

from langchain.embeddings import HuggingFaceBgeEmbeddings

288

model_name = "BAAI/bge-large-en-v1.5"

289

model_kwargs = {'device': 'cuda'}

290

encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity

291

model = HuggingFaceBgeEmbeddings(

292

model_name=model_name,

293

model_kwargs=model_kwargs,

294

encode_kwargs=encode_kwargs,

295

query_instruction="为这个句子生成表示以用于检索相关文章："

296

)

297

model.query_instruction = "为这个句子生成表示以用于检索相关文章："

298

```

299

300

301

#### Using HuggingFace Transformers

302

303

With the transformers package, you can use the model like this: First, you pass your input through the transformer model, then you select the last hidden state of the first token (i.e., [CLS]) as the sentence embedding.

304

305

```python

306

from transformers import AutoTokenizer, AutoModel

307

import torch

308

# Sentences we want sentence embeddings for

309

sentences = ["样例数据-1", "样例数据-2"]

310

311

# Load model from HuggingFace Hub

312

tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh-v1.5')

313

model = AutoModel.from_pretrained('BAAI/bge-large-zh-v1.5')

314

model.eval()

315

316

# Tokenize sentences

317

encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

318

# for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)

319

# encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')

320

321

# Compute token embeddings

322

with torch.no_grad():

323

model_output = model(**encoded_input)

324

# Perform pooling. In this case, cls pooling.

325

sentence_embeddings = model_output[0][:, 0]

326

# normalize embeddings

327

sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)

328

print("Sentence embeddings:", sentence_embeddings)

329

```

330

331

### Usage for Reranker

332

333

Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding.

334

You can get a relevance score by inputting query and passage to the reranker.

335

The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.

336

337

338

#### Using FlagEmbedding

339

```

340

pip install -U FlagEmbedding

341

```

342

343

Get relevance scores (higher scores indicate more relevance):

344

```python

345

from FlagEmbedding import FlagReranker

346

reranker = FlagReranker('BAAI/bge-reranker-large', use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation

347

348

score = reranker.compute_score(['query', 'passage'])

349

print(score)

350

351

scores = reranker.compute_score([['what is panda?', 'hi'], ['what is panda?', 'The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.']])

352

print(scores)

353

```

354

355

356

#### Using Huggingface transformers

357

358

```python

359

import torch

360

from transformers import AutoModelForSequenceClassification, AutoTokenizer

361

362

tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-large')

363

model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-large')

364

model.eval()

365

366

pairs = [['what is panda?', 'hi'], ['what is panda?', 'The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.']]

367

with torch.no_grad():

368

inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)

369

scores = model(**inputs, return_dict=True).logits.view(-1, ).float()

370

print(scores)

371

```

372

373

#### Usage reranker with the ONNX files

374

375

```python

376

from optimum.onnxruntime import ORTModelForSequenceClassification # type: ignore

377

378

import torch

379

from transformers import AutoModelForSequenceClassification, AutoTokenizer

380

381

tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-large')

382

model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base')

383

model_ort = ORTModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base', file_name="onnx/model.onnx")

384

385

# Sentences we want sentence embeddings for

386

pairs = [['what is panda?', 'hi'], ['what is panda?', 'The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.']]

387

388

# Tokenize sentences

389

encoded_input = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt')

390

391

scores_ort = model_ort(**encoded_input, return_dict=True).logits.view(-1, ).float()

392

# Compute token embeddings

393

with torch.inference_mode():

394

scores = model_ort(**encoded_input, return_dict=True).logits.view(-1, ).float()

395

396

# scores and scores_ort are identical

397

```

398

#### Usage reranker with infinity

399

400

Its also possible to deploy the onnx/torch files with the [infinity_emb](https://github.com/michaelfeil/infinity) pip package.

401

```python

402

import asyncio

403

from infinity_emb import AsyncEmbeddingEngine, EngineArgs

404

405

query='what is a panda?'

406

docs = ['The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear', "Paris is in France."]

407

408

engine = AsyncEmbeddingEngine.from_args(

409

    EngineArgs(model_name_or_path = "BAAI/bge-reranker-base", device="cpu", engine="torch" # or engine="optimum" for onnx

410

))

411

412

async def main():

413

async with engine:

414

ranking, usage = await engine.rerank(query=query, docs=docs)

415

print(list(zip(ranking, docs)))

416

asyncio.run(main())

417

```

418

419

## Evaluation

420

421

`baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**

422

For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).

423

424

- **MTEB**:

425

426

| Model Name |  Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) |  STS (10) | Summarization (1) | Classification (12) |

427

|:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|

428

| [BAAI/bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) | 1024 | 512 |  **64.23** | **54.29** |  46.08 | 87.12 | 60.03 | 83.11 | 31.61 | 75.97 |

429

| [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) |  768 | 512 | 63.55 | 53.25 |   45.77 | 86.55 | 58.86 | 82.4 | 31.07 | 75.53 |

430

| [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) |  384 | 512 | 62.17 |51.68 | 43.82 |  84.92 | 58.36 | 81.59 | 30.12 | 74.14 |

431

| [bge-large-en](https://huggingface.co/BAAI/bge-large-en) |  1024 | 512 | 63.98 |  53.9 | 46.98 | 85.8 | 59.48 | 81.56 | 32.06 | 76.21 |

432

| [bge-base-en](https://huggingface.co/BAAI/bge-base-en) |  768 | 512 |  63.36 | 53.0 | 46.32 | 85.86 | 58.7 | 81.84 | 29.27 | 75.27 |

433

| [gte-large](https://huggingface.co/thenlper/gte-large) |  1024 | 512 | 63.13 | 52.22 | 46.84 | 85.00 | 59.13 | 83.35 | 31.66 | 73.33 |

434

| [gte-base](https://huggingface.co/thenlper/gte-base) 	|  768 | 512 | 62.39 | 51.14 | 46.2 | 84.57 | 58.61 | 82.3 | 31.17 | 73.01 |

435

| [e5-large-v2](https://huggingface.co/intfloat/e5-large-v2) |  1024| 512 | 62.25 | 50.56 | 44.49 | 86.03 | 56.61 | 82.05 | 30.19 | 75.24 |

436

| [bge-small-en](https://huggingface.co/BAAI/bge-small-en) |  384 | 512 | 62.11 |  51.82 | 44.31 | 83.78 | 57.97 | 80.72 | 30.53 | 74.37 |

437

| [instructor-xl](https://huggingface.co/hkunlp/instructor-xl) |  768 | 512 | 61.79 | 49.26 | 44.74 | 86.62 | 57.29 | 83.06 | 32.32 | 61.79 |

438

| [e5-base-v2](https://huggingface.co/intfloat/e5-base-v2) |  768 | 512 | 61.5 | 50.29 | 43.80 | 85.73 | 55.91 | 81.05 | 30.28 | 73.84 |

439

| [gte-small](https://huggingface.co/thenlper/gte-small) |  384 | 512 | 61.36 | 49.46 | 44.89 | 83.54 | 57.7 | 82.07 | 30.42 | 72.31 |

440

| [text-embedding-ada-002](https://platform.openai.com/docs/guides/embeddings) | 1536 | 8192 | 60.99 | 49.25 | 45.9 | 84.89 | 56.32 | 80.97 | 30.8 | 70.93 |

441

| [e5-small-v2](https://huggingface.co/intfloat/e5-base-v2) | 384 | 512 | 59.93 | 49.04 | 39.92 | 84.67 | 54.32 | 80.39 | 31.16 | 72.94 |

442

| [sentence-t5-xxl](https://huggingface.co/sentence-transformers/sentence-t5-xxl) |  768 | 512 | 59.51 | 42.24 | 43.72 | 85.06 | 56.42 | 82.63 | 30.08 | 73.42 |

443

| [all-mpnet-base-v2](https://huggingface.co/sentence-transformers/all-mpnet-base-v2) 	|  768 | 514 	| 57.78 | 43.81 | 43.69 | 83.04 | 59.36 | 80.28 | 27.49 | 65.07 |

444

| [sgpt-bloom-7b1-msmarco](https://huggingface.co/bigscience/sgpt-bloom-7b1-msmarco) 	|  4096 | 2048 | 57.59 | 48.22 | 38.93 | 81.9 | 55.65 | 77.74 | 33.6 | 66.19 |

445

446

447

448

- **C-MTEB**:

449

We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.

450

Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.

451

452

453

|:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|

454

| [**BAAI/bge-large-zh-v1.5**](https://huggingface.co/BAAI/bge-large-zh-v1.5) | 1024 |  **64.53** | 70.46 | 56.25 | 81.6 | 69.13 | 65.84 | 48.99 |

455

| [BAAI/bge-base-zh-v1.5](https://huggingface.co/BAAI/bge-base-zh-v1.5) | 768 |  63.13 | 69.49 | 53.72 | 79.75 | 68.07 | 65.39 | 47.53 |

456

| [BAAI/bge-small-zh-v1.5](https://huggingface.co/BAAI/bge-small-zh-v1.5) | 512 | 57.82 | 61.77 | 49.11 | 70.41 | 63.96 | 60.92 | 44.18 |

457

| [BAAI/bge-large-zh](https://huggingface.co/BAAI/bge-large-zh) | 1024 | 64.20 | 71.53 | 54.98 | 78.94 | 68.32 | 65.11 | 48.39 |

458

| [bge-large-zh-noinstruct](https://huggingface.co/BAAI/bge-large-zh-noinstruct) | 1024 | 63.53 | 70.55 | 53 | 76.77 | 68.58 | 64.91 | 50.01 |

459

| [BAAI/bge-base-zh](https://huggingface.co/BAAI/bge-base-zh) | 768 | 62.96 | 69.53 | 54.12 | 77.5 | 67.07 | 64.91 | 47.63 |

460

| [multilingual-e5-large](https://huggingface.co/intfloat/multilingual-e5-large) | 1024 | 58.79 | 63.66 | 48.44 | 69.89 | 67.34 | 56.00 | 48.23 |

461

| [BAAI/bge-small-zh](https://huggingface.co/BAAI/bge-small-zh) | 512 | 58.27 |  63.07 | 49.45 | 70.35 | 63.64 | 61.48 | 45.09 |

462

| [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |

463

| [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |

464

| [multilingual-e5-base](https://huggingface.co/intfloat/multilingual-e5-base) | 768 | 55.48 | 61.63 | 46.49 | 67.07 | 65.35 | 54.35 | 40.68 |

465

| [multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small) | 384 | 55.38 | 59.95 | 45.27 | 66.45 | 65.85 | 53.86 | 45.26 |

466

| [text-embedding-ada-002(OpenAI)](https://platform.openai.com/docs/guides/embeddings/what-are-embeddings) | 1536 |  53.02 | 52.0 | 43.35 | 69.56 | 64.31 | 54.28 | 45.68 |

467

| [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 |  44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |

468

| [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 |  47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |

469

| [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |

470

471

472

- **Reranking**:

473

See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.

474

475

476

|:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|

477

| text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |

478

| multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |

479

| multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |

480

| multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |

481

| m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |

482

| m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |

483

| bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |

484

| bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |

485

| [BAAI/bge-reranker-base](https://huggingface.co/BAAI/bge-reranker-base) | 67.28 | 63.95 | 60.45 | 35.46 | 81.26 | 84.1 | 65.42 |

486

| [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) | 67.6 | 64.03 | 61.44 | 37.16 | 82.15 | 84.18 | 66.09 |

487

488

\* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval tasks

489

490

## Train

491

492

### BAAI Embedding

493

494

We pre-train the models using [retromae](https://github.com/staoxiao/RetroMAE) and train them on large-scale pairs data using contrastive learning.

495

**You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**

496

We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).

497

Note that the goal of pre-training is to reconstruct the text, and the pre-trained model cannot be used for similarity calculation directly, it needs to be fine-tuned.

498

More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).

499

500

501

502

### BGE Reranker

503

504

Cross-encoder will perform full-attention over the input pair,

505

which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.

506

Therefore, it can be used to re-rank the top-k documents returned by embedding model.

507

We train the cross-encoder on a multilingual pair data,

508

The data format is the same as embedding model, so you can fine-tune it easily following our [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/reranker).

509

More details please refer to [./FlagEmbedding/reranker/README.md](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/reranker)

510

511

512

513

## Citation

514

515

If you find this repository useful, please consider giving a star :star: and citation

516

517

```

518

@misc{bge_embedding,

519

title={C-Pack: Packaged Resources To Advance General Chinese Embedding},

520

author={Shitao Xiao and Zheng Liu and Peitian Zhang and Niklas Muennighoff},

521

year={2023},

522

eprint={2309.07597},

523

archivePrefix={arXiv},

524

primaryClass={cs.CL}

525

}

526

```

527

528

## License

529

FlagEmbedding is licensed under the [MIT License](https://github.com/FlagOpen/FlagEmbedding/blob/master/LICENSE). The released models can be used for commercial purposes free of charge.