README.md · bge-reranker-large

1

---

2

license: mit

3

language:

4

- en

5

- zh

6

tags:

7

- mteb

8

model-index:

9

- name: bge-reranker-base

10

results:

11

- task:

12

type: Reranking

13

dataset:

14

type: C-MTEB/CMedQAv1-reranking

15

name: MTEB CMedQAv1

16

config: default

17

split: test

18

revision: None

19

metrics:

20

- type: map

21

value: 81.27206722525007

22

- type: mrr

23

value: 84.14238095238095

24

- task:

25

type: Reranking

26

dataset:

27

type: C-MTEB/CMedQAv2-reranking

28

name: MTEB CMedQAv2

29

config: default

30

split: test

31

revision: None

32

metrics:

33

- type: map

34

value: 84.10369934291236

35

- type: mrr

36

value: 86.79376984126984

37

- task:

38

type: Reranking

39

dataset:

40

type: C-MTEB/Mmarco-reranking

41

name: MTEB MMarcoReranking

42

config: default

43

split: dev

44

revision: None

45

metrics:

46

- type: map

47

value: 35.4600511272538

48

- type: mrr

49

value: 34.60238095238095

50

- task:

51

type: Reranking

52

dataset:

53

type: C-MTEB/T2Reranking

54

name: MTEB T2Reranking

55

config: default

56

split: dev

57

revision: None

58

metrics:

59

- type: map

60

value: 67.27728847727172

61

- type: mrr

62

value: 77.1315192743764

63

pipeline_tag: feature-extraction

64

---

65

66

**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.**

67

68

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

69

70

71

<h4 align="center">

72

<p>

73

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

74

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

75

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

76

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

77

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

78

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

79

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

80

<p>

81

</h4>

82

83

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

84

85

86

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

87

88

89

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

90

91

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

92

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

93

- **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)

94

- **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)

95

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

96

97

## News

98

- 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.

99

- 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.

100

- 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).

101

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

102

[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:

103

- 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:

104

- 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)

105

- 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:

106

- 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)

107

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

108

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

109

- 09/12/2023: New models:

110

    - **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.

111

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

112

113

114

<details>

115

<summary>More</summary>

116

117

118

- 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.

119

- 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).

120

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

121

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

122

- 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.

123

124

</details>

125

126

127

## Model List

128

129

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

130

131

132

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

133

| [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) |  |

134

|  [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) |

135

|  [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] |   |

136

|  [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] |   |

137

|  [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: `  |

138

|  [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: `  |

139

|  [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: `  |

140

|  [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：`  |

141

|  [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：`  |

142

|  [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：`  |

143

|  [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: `  |

144

|  [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: `  |

145

|  [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: `  |

146

|  [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：`  |

147

|  [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：`  |

148

|  [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：`  |

149

150

151

[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.

152

153

[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.

154

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.

155

156

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

157

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

158

159

160

## Frequently asked questions

161

162

<details>

163

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

164

165

166

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

167

Some suggestions:

168

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

169

- 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.

170

- 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.

171

  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

172

173

174

</details>

175

176

<details>

177

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

178

179

180

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

181

182

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

183

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

184

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

185

186

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

187

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

188

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

189

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

190

191

</details>

192

193

<details>

194

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

195

196

197

198

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

199

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

200

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

201

202

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

203

it is recommended to add instructions for these short queries.

204

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

205

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

206

207

</details>

208

209

210

## Usage

211

212

### Usage for Embedding Model

213

214

Here are some examples for using `bge` models with

215

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

216

217

#### Using FlagEmbedding

218

```

219

pip install -U FlagEmbedding

220

```

221

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.

222

223

```python

224

from FlagEmbedding import FlagModel

225

sentences_1 = ["样例数据-1", "样例数据-2"]

226

sentences_2 = ["样例数据-3", "样例数据-4"]

227

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

228

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

229

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

230

embeddings_1 = model.encode(sentences_1)

231

embeddings_2 = model.encode(sentences_2)

232

similarity = embeddings_1 @ embeddings_2.T

233

print(similarity)

234

235

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

236

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

237

queries = ['query_1', 'query_2']

238

passages = ["样例文档-1", "样例文档-2"]

239

q_embeddings = model.encode_queries(queries)

240

p_embeddings = model.encode(passages)

241

scores = q_embeddings @ p_embeddings.T

242

```

243

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

244

245

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

246

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

247

248

249

#### Using Sentence-Transformers

250

251

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

252

253

```

254

pip install -U sentence-transformers

255

```

256

```python

257

from sentence_transformers import SentenceTransformer

258

sentences_1 = ["样例数据-1", "样例数据-2"]

259

sentences_2 = ["样例数据-3", "样例数据-4"]

260

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

261

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

262

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

263

similarity = embeddings_1 @ embeddings_2.T

264

print(similarity)

265

```

266

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

267

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

268

But the instruction is not needed for passages.

269

```python

270

from sentence_transformers import SentenceTransformer

271

queries = ['query_1', 'query_2']

272

passages = ["样例文档-1", "样例文档-2"]

273

instruction = "为这个句子生成表示以用于检索相关文章："

274

275

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

276

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

277

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

278

scores = q_embeddings @ p_embeddings.T

279

```

280

281

#### Using Langchain

282

283

You can use `bge` in langchain like this:

284

```python

285

from langchain.embeddings import HuggingFaceBgeEmbeddings

286

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

287

model_kwargs = {'device': 'cuda'}

288

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

289

model = HuggingFaceBgeEmbeddings(

290

model_name=model_name,

291

model_kwargs=model_kwargs,

292

encode_kwargs=encode_kwargs,

293

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

294

)

295

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

296

```

297

298

299

#### Using HuggingFace Transformers

300

301

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.

302

303

```python

304

from transformers import AutoTokenizer, AutoModel

305

import torch

306

# Sentences we want sentence embeddings for

307

sentences = ["样例数据-1", "样例数据-2"]

308

309

# Load model from HuggingFace Hub

310

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

311

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

312

model.eval()

313

314

# Tokenize sentences

315

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

316

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

317

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

318

319

# Compute token embeddings

320

with torch.no_grad():

321

model_output = model(**encoded_input)

322

# Perform pooling. In this case, cls pooling.

323

sentence_embeddings = model_output[0][:, 0]

324

# normalize embeddings

325

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

326

print("Sentence embeddings:", sentence_embeddings)

327

```

328

329

### Usage for Reranker

330

331

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

332

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

333

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

334

335

336

#### Using FlagEmbedding

337

```

338

pip install -U FlagEmbedding

339

```

340

341

Get relevance scores (higher scores indicate more relevance):

342

```python

343

from FlagEmbedding import FlagReranker

344

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

345

346

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

347

print(score)

348

349

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.']])

350

print(scores)

351

```

352

353

354

#### Using Huggingface transformers

355

356

```python

357

import torch

358

from transformers import AutoModelForSequenceClassification, AutoTokenizer

359

360

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

361

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

362

model.eval()

363

364

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.']]

365

with torch.no_grad():

366

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

367

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

368

print(scores)

369

```

370

371

#### Usage reranker with the ONNX files

372

373

```python

374

from optimum.onnxruntime import ORTModelForSequenceClassification # type: ignore

375

376

import torch

377

from transformers import AutoModelForSequenceClassification, AutoTokenizer

378

379

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

380

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

381

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

382

383

# Sentences we want sentence embeddings for

384

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.']]

385

386

# Tokenize sentences

387

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

388

389

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

390

# Compute token embeddings

391

with torch.inference_mode():

392

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

393

394

# scores and scores_ort are identical

395

```

396

#### Usage reranker with infinity

397

398

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

399

```python

400

import asyncio

401

from infinity_emb import AsyncEmbeddingEngine, EngineArgs

402

403

query='what is a panda?'

404

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

405

406

engine = AsyncEmbeddingEngine.from_args(

407

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

408

))

409

410

async def main():

411

async with engine:

412

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

413

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

414

asyncio.run(main())

415

```

416

417

## Evaluation

418

419

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

420

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

421

422

- **MTEB**:

423

424

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

425

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

426

| [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 |

427

| [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 |

428

| [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 |

429

| [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 |

430

| [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 |

431

| [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 |

432

| [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 |

433

| [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 |

434

| [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 |

435

| [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 |

436

| [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 |

437

| [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 |

438

| [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 |

439

| [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 |

440

| [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 |

441

| [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 |

442

| [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 |

443

444

445

446

- **C-MTEB**:

447

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

448

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

449

450

451

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

452

| [**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 |

453

| [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 |

454

| [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 |

455

| [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 |

456

| [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 |

457

| [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 |

458

| [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 |

459

| [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 |

460

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

461

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

462

| [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 |

463

| [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 |

464

| [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 |

465

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

466

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

467

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

468

469

470

- **Reranking**:

471

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

472

473

474

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

475

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

476

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

477

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

478

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

479

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

480

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

481

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

482

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

483

| [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 |

484

| [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 |

485

486

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

487

488

## Train

489

490

### BAAI Embedding

491

492

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

493

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

494

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

495

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.

496

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

497

498

499

500

### BGE Reranker

501

502

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

503

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

504

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

505

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

506

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).

507

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

508

509

510

511

## Citation

512

513

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

514

515

```

516

@misc{bge_embedding,

517

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

518

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

519

year={2023},

520

eprint={2309.07597},

521

archivePrefix={arXiv},

522

primaryClass={cs.CL}

523

}

524

```

525

526

## License

527

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.