ldwang commited on
Commit
60e9d3d
1 Parent(s): 2808477
Files changed (1) hide show
  1. README.md +370 -0
README.md CHANGED
@@ -1,3 +1,373 @@
1
  ---
2
  license: mit
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
  ---
4
+
5
+
6
+ <h1 align="center">FlagEmbedding</h1>
7
+
8
+
9
+ <h4 align="center">
10
+ <p>
11
+ <a href=#model-list>Model List</a> |
12
+ <a href=#usage>Usage</a> |
13
+ <a href="#evaluation">Evaluation</a> |
14
+ <a href="#train">Train</a> |
15
+ <a href="#contact">Contact</a> |
16
+ <a href="#license">License</a>
17
+ <p>
18
+ </h4>
19
+
20
+ More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
21
+
22
+
23
+
24
+ [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
25
+
26
+ FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
27
+ And it also can be used in vector databases for LLMs.
28
+
29
+ ************* 🌟**Updates**🌟 *************
30
+ - 09/12/2023: New Release:
31
+ - **New reranker model**: release a cross-encoder model bge-reranker-base, which is more powerful than embedding model. We recommend to use/fine-tune it to re-rank top-k documents returned by embedding models.
32
+ - **update embedding model**: release bge-*-v1.5 embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.
33
+ - 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.
34
+ - 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).
35
+ - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
36
+ - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:
37
+ - 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.
38
+
39
+
40
+ ## Model List
41
+
42
+ `bge` is short for `BAAI general embedding`.
43
+
44
+ | Model | Language | | Description | query instruction for retrieval\* |
45
+ |:-------------------------------|:--------:| :--------:| :--------:|:--------:|
46
+ | [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 \** | |
47
+ | [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 \** | |
48
+ | [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: ` |
49
+ | [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: ` |
50
+ | [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: ` |
51
+ | [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:` |
52
+ | [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:` |
53
+ | [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:` |
54
+ | [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: ` |
55
+ | [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: ` |
56
+ | [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: ` |
57
+ | [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:` |
58
+ | [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:` |
59
+ | [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:` |
60
+
61
+
62
+ \*: 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.
63
+
64
+ \**: To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models.
65
+ 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.
66
+
67
+
68
+ ## Frequently asked questions
69
+
70
+ <details>
71
+ <summary>1. How to fine-tune bge embedding model?</summary>
72
+
73
+ <!-- ### How to fine-tune bge embedding model? -->
74
+ Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.
75
+ Some suggestions:
76
+ - Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#data-format), which can improve the retrieval performance.
77
+ - 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.
78
+ - 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. Hard negatives also are needed to fine-tune reranker.
79
+
80
+
81
+ </details>
82
+
83
+ <details>
84
+ <summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>
85
+
86
+ <!-- ### The similarity score between two dissimilar sentences is higher than 0.5 -->
87
+ **Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**
88
+
89
+ Since we finetune the models by contrastive learning with a temperature of 0.01,
90
+ the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].
91
+ So a similarity score greater than 0.5 does not indicate that the two sentences are similar.
92
+
93
+ For downstream tasks, such as passage retrieval or semantic similarity,
94
+ **what matters is the relative order of the scores, not the absolute value.**
95
+ If you need to filter similar sentences based on a similarity threshold,
96
+ please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).
97
+
98
+ </details>
99
+
100
+ <details>
101
+ <summary>3. When does the query instruction need to be used</summary>
102
+
103
+ <!-- ### When does the query instruction need to be used -->
104
+
105
+ For a retrieval task that uses short queries to find long related documents,
106
+ it is recommended to add instructions for these short queries.
107
+ **The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**
108
+ In all cases, the documents/passages do not need to add the instruction.
109
+
110
+ </details>
111
+
112
+
113
+ ## Usage
114
+
115
+ ### Usage for Embedding Model
116
+
117
+ Here are some examples for using `bge` models with
118
+ [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
119
+
120
+ #### Using FlagEmbedding
121
+ ```
122
+ pip install -U FlagEmbedding
123
+ ```
124
+ 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.
125
+
126
+ ```python
127
+ from FlagEmbedding import FlagModel
128
+ sentences_1 = ["样例数据-1", "样例数据-2"]
129
+ sentences_2 = ["样例数据-3", "样例数据-4"]
130
+ model = FlagModel('BAAI/bge-large-zh', query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章:")
131
+ embeddings_1 = model.encode(sentences_1)
132
+ embeddings_2 = model.encode(sentences_2)
133
+ similarity = embeddings_1 @ embeddings_2.T
134
+ print(similarity)
135
+
136
+ # for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query
137
+ # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
138
+ queries = ['query_1', 'query_2']
139
+ passages = ["样例文档-1", "样例文档-2"]
140
+ q_embeddings = model.encode_queries(queries)
141
+ p_embeddings = model.encode(passages)
142
+ scores = q_embeddings @ p_embeddings.T
143
+ ```
144
+ For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
145
+
146
+ By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.
147
+ You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.
148
+
149
+
150
+ #### Using Sentence-Transformers
151
+
152
+ You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):
153
+
154
+ ```
155
+ pip install -U sentence-transformers
156
+ ```
157
+ ```python
158
+ from sentence_transformers import SentenceTransformer
159
+ sentences_1 = ["样例数据-1", "样例数据-2"]
160
+ sentences_2 = ["样例数据-3", "样例数据-4"]
161
+ model = SentenceTransformer('BAAI/bge-large-zh')
162
+ embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
163
+ embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
164
+ similarity = embeddings_1 @ embeddings_2.T
165
+ print(similarity)
166
+ ```
167
+ For s2p(short query to long passage) retrieval task,
168
+ each short query should start with an instruction (instructions see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list)).
169
+ But the instruction is not needed for passages.
170
+ ```python
171
+ from sentence_transformers import SentenceTransformer
172
+ queries = ['query_1', 'query_2']
173
+ passages = ["样例文档-1", "样例文档-2"]
174
+ instruction = "为这个句子生成表示以用于检索相关文章:"
175
+
176
+ model = SentenceTransformer('BAAI/bge-large-zh')
177
+ q_embeddings = model.encode([instruction+q for q in queries], normalize_embeddings=True)
178
+ p_embeddings = model.encode(passages, normalize_embeddings=True)
179
+ scores = q_embeddings @ p_embeddings.T
180
+ ```
181
+
182
+ #### Using Langchain
183
+
184
+ You can use `bge` in langchain like this:
185
+ ```python
186
+ from langchain.embeddings import HuggingFaceBgeEmbeddings
187
+ model_name = "BAAI/bge-small-en"
188
+ model_kwargs = {'device': 'cuda'}
189
+ encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
190
+ model = HuggingFaceBgeEmbeddings(
191
+ model_name=model_name,
192
+ model_kwargs=model_kwargs,
193
+ encode_kwargs=encode_kwargs,
194
+ query_instruction="为这个句子生成表示以用于检索相关文章:"
195
+ )
196
+ model.query_instruction = "为这个句子生成表示以用于检索相关文章:"
197
+ ```
198
+
199
+
200
+ #### Using HuggingFace Transformers
201
+
202
+ 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.
203
+
204
+ ```python
205
+ from transformers import AutoTokenizer, AutoModel
206
+ import torch
207
+ # Sentences we want sentence embeddings for
208
+ sentences = ["样例数据-1", "样例数据-2"]
209
+
210
+ # Load model from HuggingFace Hub
211
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
212
+ model = AutoModel.from_pretrained('BAAI/bge-large-zh')
213
+ model.eval()
214
+
215
+ # Tokenize sentences
216
+ encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
217
+ # for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)
218
+ # encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')
219
+
220
+ # Compute token embeddings
221
+ with torch.no_grad():
222
+ model_output = model(**encoded_input)
223
+ # Perform pooling. In this case, cls pooling.
224
+ sentence_embeddings = model_output[0][:, 0]
225
+ # normalize embeddings
226
+ sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
227
+ print("Sentence embeddings:", sentence_embeddings)
228
+ ```
229
+
230
+ ### Usage for Reranker
231
+
232
+ You can get a relevance score by inputting query and passage to the reranker.
233
+ The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.
234
+
235
+
236
+ #### Using FlagEmbedding
237
+ ```
238
+ pip install -U FlagEmbedding
239
+ ```
240
+
241
+ Get relevance score:
242
+ ```python
243
+ from FlagEmbedding import FlagReranker
244
+ reranker = FlagReranker('BAAI/bge-reranker-base', use_fp16=True) #use fp16 can speed up computing
245
+
246
+ score = reranker.compute_score(['query', 'passage'])
247
+ print(score)
248
+
249
+ 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.']])
250
+ print(scores)
251
+ ```
252
+
253
+
254
+ #### Using Huggingface transformers
255
+
256
+ ```python
257
+ import torch
258
+ from transformers import AutoModelForSequenceClassification, AutoTokenizer, BatchEncoding, PreTrainedTokenizerFast
259
+
260
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-base')
261
+ model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base')
262
+ model.eval()
263
+
264
+ 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.']]
265
+ with torch.no_grad():
266
+ inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
267
+ scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
268
+ print(scores)
269
+ ```
270
+
271
+ ## Evaluation
272
+
273
+ `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
274
+ For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
275
+
276
+ - **MTEB**:
277
+
278
+ | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
279
+ |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
280
+ | [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 |
281
+ | [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 |
282
+ | [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 |
283
+ | [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 |
284
+ | [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 |
285
+ | [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 |
286
+ | [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 |
287
+ | [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 |
288
+ | [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 |
289
+ | [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 |
290
+ | [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 |
291
+ | [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 |
292
+ | [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 |
293
+ | [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 |
294
+ | [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 |
295
+ | [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 |
296
+ | [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 |
297
+
298
+
299
+
300
+ - **C-MTEB**:
301
+ We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
302
+ Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
303
+
304
+ | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
305
+ |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
306
+ | [**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 |
307
+ | [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 |
308
+ | [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 |
309
+ | [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 |
310
+ | [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 |
311
+ | [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 |
312
+ | [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 |
313
+ | [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 |
314
+ | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |
315
+ | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |
316
+ | [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 |
317
+ | [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 |
318
+ | [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 |
319
+ | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |
320
+ | [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |
321
+ | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |
322
+
323
+
324
+ - **Reranking**:
325
+ See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.
326
+
327
+ | Model | T2Reranking | T2RerankingZh2En\* | T2RerankingEn2Zh\* | MmarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
328
+ |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
329
+ | text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |
330
+ | multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |
331
+ | multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |
332
+ | multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |
333
+ | m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |
334
+ | m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |
335
+ | bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |
336
+ | bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |
337
+ | [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 |
338
+ | [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 |
339
+
340
+ \* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval task
341
+
342
+ ## Train
343
+
344
+ ### BAAI Embedding
345
+
346
+ We pre-train the models using retromae and train them on large-scale pairs data using contrastive learning.
347
+ **You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**
348
+ We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).
349
+ 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.
350
+ More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
351
+
352
+
353
+
354
+ ### BGE Reranker
355
+
356
+ Cross-encoder will perform full-attention over the input pair,
357
+ which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.
358
+ Therefore, it can be used to re-rank the top-k documents returned by embedding model.
359
+ We train the cross-encoder on a multilingual pair data,
360
+ The data format is the same as embedding model, so you can fine-tune it easily following our example.
361
+ More details pelease refer to [./FlagEmbedding/reranker/README.md](./FlagEmbedding/reranker/README.md)
362
+
363
+
364
+ ## Contact
365
+ If you have any question or suggestion related to this project, feel free to open an issue or pull request.
366
+ You also can email Shitao Xiao([email protected]) and Zheng Liu([email protected]).
367
+
368
+
369
+ ## License
370
+ 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.
371
+
372
+
373
+