README.md
27.0 KB · 412 lines · markdown Raw
1 ---
2 license: mit
3 language:
4 - zh
5 ---
6
7
8 <h1 align="center">FlagEmbedding</h1>
9
10
11 <h4 align="center">
12 <p>
13 <a href=#model-list>Model List</a> |
14 <a href=#frequently-asked-questions>FAQ</a> |
15 <a href=#usage>Usage</a> |
16 <a href="#evaluation">Evaluation</a> |
17 <a href="#train">Train</a> |
18 <a href="#contact">Contact</a> |
19 <a href="#citation">Citation</a> |
20 <a href="#license">License</a>
21 <p>
22 </h4>
23
24 More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
25
26
27 [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
28
29 FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
30 And it also can be used in vector databases for LLMs.
31
32 ************* 🌟**Updates**🌟 *************
33 - 10/12/2023: Release [LLM-Embedder](./FlagEmbedding/llm_embedder/README.md), a unified embedding model to support diverse retrieval augmentation needs for LLMs. [Paper](https://arxiv.org/pdf/2310.07554.pdf) :fire:
34 - 09/15/2023: The [technical report](https://arxiv.org/pdf/2309.07597.pdf) of BGE has been released
35 - 09/15/2023: The [masive training data](https://data.baai.ac.cn/details/BAAI-MTP) of BGE has been released
36 - 09/12/2023: New models:
37 - **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.
38 - **update embedding model**: release `bge-*-v1.5` embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.
39
40
41 <details>
42 <summary>More</summary>
43 <!-- ### More -->
44
45 - 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.
46 - 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).
47 - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
48 - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:
49 - 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.
50
51 </details>
52
53
54 ## Model List
55
56 `bge` is short for `BAAI general embedding`.
57
58 | Model | Language | | Description | query instruction for retrieval [1] |
59 |:-------------------------------|:--------:| :--------:| :--------:|:--------:|
60 | [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) |
61 | [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] | |
62 | [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] | |
63 | [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: ` |
64 | [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: ` |
65 | [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: ` |
66 | [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：` |
67 | [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：` |
68 | [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：` |
69 | [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: ` |
70 | [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: ` |
71 | [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: ` |
72 | [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：` |
73 | [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：` |
74 | [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：` |
75
76
77 [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.
78
79 [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.
80 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.
81
82 All models have been uploaded to Huggingface Hub, and you can see them at https://huggingface.co/BAAI.
83 If you cannot open the Huggingface Hub, you also can download the models at https://model.baai.ac.cn/models .
84
85
86 ## Frequently asked questions
87
88 <details>
89 <summary>1. How to fine-tune bge embedding model?</summary>
90
91 <!-- ### How to fine-tune bge embedding model? -->
92 Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.
93 Some suggestions:
94 - Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#hard-negatives), which can improve the retrieval performance.
95 - 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.
96 - 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.
97
98
99 </details>
100
101 <details>
102 <summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>
103
104 <!-- ### The similarity score between two dissimilar sentences is higher than 0.5 -->
105 **Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**
106
107 Since we finetune the models by contrastive learning with a temperature of 0.01,
108 the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].
109 So a similarity score greater than 0.5 does not indicate that the two sentences are similar.
110
111 For downstream tasks, such as passage retrieval or semantic similarity,
112 **what matters is the relative order of the scores, not the absolute value.**
113 If you need to filter similar sentences based on a similarity threshold,
114 please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).
115
116 </details>
117
118 <details>
119 <summary>3. When does the query instruction need to be used</summary>
120
121 <!-- ### When does the query instruction need to be used -->
122
123 For the `bge-*-v1.5`, we improve its retrieval ability when not using instruction.
124 No instruction only has a slight degradation in retrieval performance compared with using instruction.
125 So you can generate embedding without instruction in all cases for convenience.
126
127 For a retrieval task that uses short queries to find long related documents,
128 it is recommended to add instructions for these short queries.
129 **The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**
130 In all cases, the documents/passages do not need to add the instruction.
131
132 </details>
133
134
135 ## Usage
136
137 ### Usage for Embedding Model
138
139 Here are some examples for using `bge` models with
140 [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
141
142 #### Using FlagEmbedding
143 ```
144 pip install -U FlagEmbedding
145 ```
146 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.
147
148 ```python
149 from FlagEmbedding import FlagModel
150 sentences_1 = ["样例数据-1", "样例数据-2"]
151 sentences_2 = ["样例数据-3", "样例数据-4"]
152 model = FlagModel('BAAI/bge-large-zh-v1.5',
153 query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章：",
154 use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation
155 embeddings_1 = model.encode(sentences_1)
156 embeddings_2 = model.encode(sentences_2)
157 similarity = embeddings_1 @ embeddings_2.T
158 print(similarity)
159
160 # for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query
161 # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
162 queries = ['query_1', 'query_2']
163 passages = ["样例文档-1", "样例文档-2"]
164 q_embeddings = model.encode_queries(queries)
165 p_embeddings = model.encode(passages)
166 scores = q_embeddings @ p_embeddings.T
167 ```
168 For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
169
170 By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.
171 You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.
172
173
174 #### Using Sentence-Transformers
175
176 You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):
177
178 ```
179 pip install -U sentence-transformers
180 ```
181 ```python
182 from sentence_transformers import SentenceTransformer
183 sentences_1 = ["样例数据-1", "样例数据-2"]
184 sentences_2 = ["样例数据-3", "样例数据-4"]
185 model = SentenceTransformer('BAAI/bge-large-zh-v1.5')
186 embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
187 embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
188 similarity = embeddings_1 @ embeddings_2.T
189 print(similarity)
190 ```
191 For s2p(short query to long passage) retrieval task,
192 each short query should start with an instruction (instructions see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list)).
193 But the instruction is not needed for passages.
194 ```python
195 from sentence_transformers import SentenceTransformer
196 queries = ['query_1', 'query_2']
197 passages = ["样例文档-1", "样例文档-2"]
198 instruction = "为这个句子生成表示以用于检索相关文章："
199
200 model = SentenceTransformer('BAAI/bge-large-zh-v1.5')
201 q_embeddings = model.encode([instruction+q for q in queries], normalize_embeddings=True)
202 p_embeddings = model.encode(passages, normalize_embeddings=True)
203 scores = q_embeddings @ p_embeddings.T
204 ```
205
206 #### Using Langchain
207
208 You can use `bge` in langchain like this:
209 ```python
210 from langchain.embeddings import HuggingFaceBgeEmbeddings
211 model_name = "BAAI/bge-large-en-v1.5"
212 model_kwargs = {'device': 'cuda'}
213 encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
214 model = HuggingFaceBgeEmbeddings(
215 model_name=model_name,
216 model_kwargs=model_kwargs,
217 encode_kwargs=encode_kwargs,
218 query_instruction="为这个句子生成表示以用于检索相关文章："
219 )
220 model.query_instruction = "为这个句子生成表示以用于检索相关文章："
221 ```
222
223
224 #### Using HuggingFace Transformers
225
226 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.
227
228 ```python
229 from transformers import AutoTokenizer, AutoModel
230 import torch
231 # Sentences we want sentence embeddings for
232 sentences = ["样例数据-1", "样例数据-2"]
233
234 # Load model from HuggingFace Hub
235 tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh-v1.5')
236 model = AutoModel.from_pretrained('BAAI/bge-large-zh-v1.5')
237 model.eval()
238
239 # Tokenize sentences
240 encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
241 # for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)
242 # encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')
243
244 # Compute token embeddings
245 with torch.no_grad():
246 model_output = model(**encoded_input)
247 # Perform pooling. In this case, cls pooling.
248 sentence_embeddings = model_output[0][:, 0]
249 # normalize embeddings
250 sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
251 print("Sentence embeddings:", sentence_embeddings)
252 ```
253
254 ### Usage for Reranker
255
256 Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding.
257 You can get a relevance score by inputting query and passage to the reranker.
258 The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.
259
260
261 #### Using FlagEmbedding
262 ```
263 pip install -U FlagEmbedding
264 ```
265
266 Get relevance scores (higher scores indicate more relevance):
267 ```python
268 from FlagEmbedding import FlagReranker
269 reranker = FlagReranker('BAAI/bge-reranker-large', use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation
270
271 score = reranker.compute_score(['query', 'passage'])
272 print(score)
273
274 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.']])
275 print(scores)
276 ```
277
278
279 #### Using Huggingface transformers
280
281 ```python
282 import torch
283 from transformers import AutoModelForSequenceClassification, AutoTokenizer
284
285 tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-large')
286 model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-large')
287 model.eval()
288
289 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.']]
290 with torch.no_grad():
291 inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
292 scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
293 print(scores)
294 ```
295
296 ## Evaluation
297
298 `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
299 For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
300
301 - **MTEB**:
302
303 | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
304 |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
305 | [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 |
306 | [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 |
307 | [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 |
308 | [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 |
309 | [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 |
310 | [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 |
311 | [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 |
312 | [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 |
313 | [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 |
314 | [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 |
315 | [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 |
316 | [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 |
317 | [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 |
318 | [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 |
319 | [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 |
320 | [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 |
321 | [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 |
322
323
324
325 - **C-MTEB**:
326 We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
327 Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
328
329 | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
330 |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
331 | [**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 |
332 | [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 |
333 | [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 |
334 | [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 |
335 | [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 |
336 | [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 |
337 | [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 |
338 | [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 |
339 | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |
340 | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |
341 | [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 |
342 | [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 |
343 | [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 |
344 | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |
345 | [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |
346 | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |
347
348
349 - **Reranking**:
350 See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.
351
352 | Model | T2Reranking | T2RerankingZh2En\* | T2RerankingEn2Zh\* | MMarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
353 |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
354 | text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |
355 | multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |
356 | multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |
357 | multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |
358 | m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |
359 | m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |
360 | bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |
361 | bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |
362 | [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 |
363 | [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 |
364
365 \* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval tasks
366
367 ## Train
368
369 ### BAAI Embedding
370
371 We pre-train the models using [retromae](https://github.com/staoxiao/RetroMAE) and train them on large-scale pairs data using contrastive learning.
372 **You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**
373 We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).
374 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.
375 More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
376
377
378
379 ### BGE Reranker
380
381 Cross-encoder will perform full-attention over the input pair,
382 which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.
383 Therefore, it can be used to re-rank the top-k documents returned by embedding model.
384 We train the cross-encoder on a multilingual pair data,
385 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).
386 More details please refer to [./FlagEmbedding/reranker/README.md](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/reranker)
387
388
389 ## Contact
390 If you have any question or suggestion related to this project, feel free to open an issue or pull request.
391 You also can email Shitao Xiao(stxiao@baai.ac.cn) and Zheng Liu(liuzheng@baai.ac.cn).
392
393
394 ## Citation
395
396 If you find this repository useful, please consider giving a star :star: and citation
397
398 ```
399 @misc{bge_embedding,
400 title={C-Pack: Packaged Resources To Advance General Chinese Embedding},
401 author={Shitao Xiao and Zheng Liu and Peitian Zhang and Niklas Muennighoff},
402 year={2023},
403 eprint={2309.07597},
404 archivePrefix={arXiv},
405 primaryClass={cs.CL}
406 }
407 ```
408
409 ## License
410 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.
411
412