ldwang commited on
Commit
c9c382e
1 Parent(s): 4bf3c54
Files changed (1) hide show
  1. README.md +201 -92
README.md CHANGED
@@ -7,7 +7,7 @@ tags:
7
  - transformers
8
  license: mit
9
  language:
10
- - zh
11
  ---
12
 
13
 
@@ -27,15 +27,21 @@ language:
27
 
28
  More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
29
 
 
 
30
  [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
31
 
32
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
33
- And it also can be used in vector database for LLMs.
34
 
35
  ************* 🌟**Updates**🌟 *************
36
- - 08/09/2023: BGE Models are integrated into **Langchain**, you can use it like [**this**](#using-langchain); C-MTEB **leaderboard** is [avaliable](https://huggingface.co/spaces/mteb/leaderboard).
 
 
 
 
37
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
38
- - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!**
39
  - 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.
40
 
41
 
@@ -43,21 +49,80 @@ And it also can be used in vector database for LLMs.
43
 
44
  `bge` is short for `BAAI general embedding`.
45
 
46
- | Model | Language | Description | query instruction for retrieval\* |
47
- |:-------------------------------|:--------:| :--------:| :--------:|
48
- | [BAAI/bge-large-en](https://huggingface.co/BAAI/bge-large-en) | English | rank **1st** in [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard | `Represent this sentence for searching relevant passages: ` |
49
- | [BAAI/bge-base-en](https://huggingface.co/BAAI/bge-base-en) | English | rank **2nd** in [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard | `Represent this sentence for searching relevant passages: ` |
50
- | [BAAI/bge-small-en](https://huggingface.co/BAAI/bge-small-en) | English | a small-scale model but with competitive performance | `Represent this sentence for searching relevant passages: ` |
51
- | [BAAI/bge-large-zh](https://huggingface.co/BAAI/bge-large-zh) | Chinese | 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:` |
52
- | [BAAI/bge-large-zh-noinstruct](https://huggingface.co/BAAI/bge-large-zh-noinstruct) | Chinese | This model is trained without instruction, and rank **2nd** in [C-MTEB](https://github.com/FlagOpen/FlagEmbedding/tree/master/C_MTEB) benchmark | |
53
- | [BAAI/bge-base-zh](https://huggingface.co/BAAI/bge-base-zh) | Chinese | a base-scale model but has similar ability with `bge-large-zh` | `为这个句子生成表示以用于检索相关文章:` |
54
- | [BAAI/bge-small-zh](https://huggingface.co/BAAI/bge-small-zh) | Chinese | 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:` |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55
 
56
- \*: If you need to search the **long** relevant passages to a **short** query (s2p retrieval task), you need 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** need to be added to passages.
57
 
58
  ## Usage
59
 
60
- Here are some examples to use `bge` models with
 
 
61
  [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
62
 
63
  #### Using FlagEmbedding
@@ -68,14 +133,15 @@ If it doesn't work for you, you can see [FlagEmbedding](https://github.com/FlagO
68
 
69
  ```python
70
  from FlagEmbedding import FlagModel
71
- sentences = ["样例数据-1", "样例数据-2"]
 
72
  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:")
73
- embeddings_1 = model.encode(sentences)
74
- embeddings_2 = model.encode(sentences)
75
  similarity = embeddings_1 @ embeddings_2.T
76
  print(similarity)
77
 
78
- # for s2p(short query to long passage) retrieval task, please use encode_queries() which will automatically add the instruction to each query
79
  # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
80
  queries = ['query_1', 'query_2']
81
  passages = ["样例文档-1", "样例文档-2"]
@@ -83,24 +149,26 @@ q_embeddings = model.encode_queries(queries)
83
  p_embeddings = model.encode(passages)
84
  scores = q_embeddings @ p_embeddings.T
85
  ```
86
- The value of argument `query_instruction_for_retrieval` see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
87
 
88
- FlagModel will use all available GPUs when encoding, please set `os.environ["CUDA_VISIBLE_DEVICES"]` to choose GPU.
 
89
 
90
 
91
  #### Using Sentence-Transformers
92
 
93
- Using this model also is easy when you have [sentence-transformers](https://www.SBERT.net) installed:
94
 
95
  ```
96
  pip install -U sentence-transformers
97
  ```
98
  ```python
99
  from sentence_transformers import SentenceTransformer
100
- sentences = ["样例数据-1", "样例数据-2"]
 
101
  model = SentenceTransformer('BAAI/bge-large-zh')
102
- embeddings_1 = model.encode(sentences, normalize_embeddings=True)
103
- embeddings_2 = model.encode(sentences, normalize_embeddings=True)
104
  similarity = embeddings_1 @ embeddings_2.T
105
  print(similarity)
106
  ```
@@ -127,17 +195,19 @@ from langchain.embeddings import HuggingFaceBgeEmbeddings
127
  model_name = "BAAI/bge-small-en"
128
  model_kwargs = {'device': 'cuda'}
129
  encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
130
- model_norm = HuggingFaceBgeEmbeddings(
131
  model_name=model_name,
132
  model_kwargs=model_kwargs,
133
- encode_kwargs=encode_kwargs
 
134
  )
 
135
  ```
136
 
137
 
138
  #### Using HuggingFace Transformers
139
 
140
- With 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 first token (i.e., [CLS]) as the sentence embedding.
141
 
142
  ```python
143
  from transformers import AutoTokenizer, AutoModel
@@ -148,6 +218,7 @@ sentences = ["样例数据-1", "样例数据-2"]
148
  # Load model from HuggingFace Hub
149
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
150
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
 
151
 
152
  # Tokenize sentences
153
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
@@ -164,21 +235,65 @@ sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, di
164
  print("Sentence embeddings:", sentence_embeddings)
165
  ```
166
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
167
 
168
  ## Evaluation
 
169
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
170
- More details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
171
 
172
  - **MTEB**:
173
 
174
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
175
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
176
- | [**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** |
177
- | [**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 |
 
 
 
178
  | [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 |
179
  | [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 |
180
  | [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 |
181
- | [**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 |
182
  | [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 |
183
  | [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 |
184
  | [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 |
@@ -187,86 +302,80 @@ More details and evaluation tools see our [scripts](https://github.com/FlagOpen/
187
  | [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 |
188
  | [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 |
189
  | [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 |
190
- | [all-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2) | 384 | 512 | 56.53 | 42.69 | 41.81 | 82.41 | 58.44 | 79.8 | 27.9 | 63.21 |
191
- | [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) | 384 | 512 | 56.26 | 41.95 | 42.35 | 82.37 | 58.04 | 78.9 | 30.81 | 63.05 |
192
- | [contriever-base-msmarco](https://huggingface.co/nthakur/contriever-base-msmarco) | 768 | 512 | 56.00 | 41.88 | 41.1 | 82.54 | 53.14 | 76.51 | 30.36 | 66.68 |
193
- | [sentence-t5-base](https://huggingface.co/sentence-transformers/sentence-t5-base) | 768 | 512 | 55.27 | 33.63 | 40.21 | 85.18 | 53.09 | 81.14 | 31.39 | 69.81 |
194
 
195
 
196
 
197
  - **C-MTEB**:
198
- We create a benchmark C-MTEB for chinese text embedding which consists of 31 datasets from 6 tasks.
199
  Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
200
 
201
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
202
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
203
- | [**bge-large-zh**](https://huggingface.co/BAAI/bge-large-zh) | 1024 | **64.20** | **71.53** | **53.23** | **78.94** | 72.26 | **65.11** | 48.39 |
204
- | [**bge-large-zh-noinstruct**](https://huggingface.co/BAAI/bge-large-zh-noinstruct) | 1024 | 63.53 | 70.55 | 50.98 | 76.77 | **72.49** | 64.91 | **50.01** |
205
- | [**BAAI/bge-base-zh**](https://huggingface.co/BAAI/bge-base-zh) | 768 | 62.96 | 69.53 | 52.05 | 77.5 | 70.98 | 64.91 | 47.63 |
206
- | [**BAAI/bge-small-zh**](https://huggingface.co/BAAI/bge-small-zh) | 512 | 58.27 | 63.07 | 46.87 | 70.35 | 67.78 | 61.48 | 45.09 |
207
- | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 |56.91 | 48.15 | 63.99 | 70.28 | 59.34 | 47.68 |
208
- | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 |54.75 | 48.64 | 64.3 | 71.22 | 59.66 | 48.88 |
209
- | [text-embedding-ada-002(OpenAI)](https://platform.openai.com/docs/guides/embeddings/what-are-embeddings) | 1536 | 53.02 | 52.0 | 40.61 | 69.56 | 67.38 | 54.28 | 45.68 |
210
- | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 39.41 | 66.62 | 65.29 | 49.25 | 44.39 |
211
- | [text2vec](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 41.71 | 67.41 | 65.18 | 49.45 | 37.66 |
212
- | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 41.98 | 70.86 | 63.42 | 49.16 | 30.02 |
213
-
214
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215
 
216
  ## Train
217
- This section will introduce the way we used to train the general embedding.
218
- The training scripts are in [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md),
219
- and we provide some examples to do [pre-train](https://github.com/FlagOpen/FlagEmbedding/blob/master/examples/pretrain/README.md) and [fine-tune](https://github.com/FlagOpen/FlagEmbedding/blob/master/examples/finetune/README.md).
220
-
221
-
222
- **1. RetroMAE Pre-train**
223
- We pre-train the model following the method [retromae](https://github.com/staoxiao/RetroMAE),
224
- which shows promising improvement in retrieval task ([paper](https://aclanthology.org/2022.emnlp-main.35.pdf)).
225
- The pre-training was conducted on 24 A100(40G) GPUs with a batch size of 720.
226
- In retromae, the mask ratio of encoder and decoder are 0.3, 0.5 respectively.
227
- We used the AdamW optimizer and the learning rate is 2e-5.
228
 
229
- **Pre-training data**:
230
- - English:
231
- - [Pile](https://pile.eleuther.ai/)
232
- - [wikipedia](https://huggingface.co/datasets/wikipedia)
233
- - [msmarco](https://huggingface.co/datasets/Tevatron/msmarco-passage-corpus)
234
- - Chinese:
235
- - [wudao](https://github.com/BAAI-WuDao/Data)
236
 
 
 
 
 
 
237
 
238
- **2. Finetune**
239
- We fine-tune the model using a contrastive objective.
240
- The format of input data is a triple`(query, positive, negative)`.
241
- Besides the negative in the triple, we also adopt in-batch negatives strategy.
242
- We employ the cross-device negatives sharing method to share negatives among different GPUs,
243
- which can dramatically **increase the number of negatives**.
244
 
245
- We trained our model on 48 A100(40G) GPUs with a large batch size of 32,768 (so there are **65,535** negatives for each query in a batch).
246
- We used the AdamW optimizer and the learning rate is 1e-5.
247
- The temperature for contrastive loss is 0.01.
248
 
249
- Besides, we add instruction to the query for s2p(short query to long passage) retrieval task in the training (add nothing to passages).
250
- For English, the instruction is `Represent this sentence for searching relevant passages: `;
251
- For Chinese, the instruction is `为这个句子生成表示以用于检索相关文章:`.
252
- In the evaluation, the instruction should be added for queries in retrieval task, not be added for other tasks.
253
- Noted that the instruction is not needed for passages.
254
 
255
- The finetune script is accessible in this repository: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
256
- You can easily finetune your model with it.
 
 
 
 
257
 
258
- **Training data**:
259
 
260
- - For English, we collect 230M text pairs from [wikipedia](https://huggingface.co/datasets/wikipedia), [cc-net](https://github.com/facebookresearch/cc_net), and so on.
 
 
261
 
262
- - For chinese, we collect 120M text pairs from [wudao](https://github.com/BAAI-WuDao/Data), [simclue](https://github.com/CLUEbenchmark/SimCLUE) and so on.
263
 
264
- **The data collection is to be released in the future.**
265
-
266
- We will continually update the embedding models and training codes,
267
- hoping to promote the development of the embedding model community.
268
 
269
 
270
 
271
- ## License
272
- FlagEmbedding is licensed under [MIT License](https://github.com/FlagOpen/FlagEmbedding/blob/master/LICENSE). The released models can be used for commercial purposes free of charge.
 
7
  - transformers
8
  license: mit
9
  language:
10
+ - en
11
  ---
12
 
13
 
 
27
 
28
  More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
29
 
30
+
31
+
32
  [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
33
 
34
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
35
+ And it also can be used in vector databases for LLMs.
36
 
37
  ************* 🌟**Updates**🌟 *************
38
+ - 09/12/2023: New Release:
39
+ - **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.
40
+ - **update embedding model**: release bge-*-v1.5 embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.
41
+ - 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.
42
+ - 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).
43
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
44
+ - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:
45
  - 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.
46
 
47
 
 
49
 
50
  `bge` is short for `BAAI general embedding`.
51
 
52
+ | Model | Language | | Description | query instruction for retrieval\* |
53
+ |:-------------------------------|:--------:| :--------:| :--------:|:--------:|
54
+ | [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 \** | |
55
+ | [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 \** | |
56
+ | [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: ` |
57
+ | [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: ` |
58
+ | [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: ` |
59
+ | [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:` |
60
+ | [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:` |
61
+ | [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:` |
62
+ | [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: ` |
63
+ | [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: ` |
64
+ | [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: ` |
65
+ | [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:` |
66
+ | [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:` |
67
+ | [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:` |
68
+
69
+
70
+ \*: 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.
71
+
72
+ \**: To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models.
73
+ 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.
74
+
75
+
76
+ ## Frequently asked questions
77
+
78
+ <details>
79
+ <summary>1. How to fine-tune bge embedding model?</summary>
80
+
81
+ <!-- ### How to fine-tune bge embedding model? -->
82
+ Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.
83
+ Some suggestions:
84
+ - Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#data-format), which can improve the retrieval performance.
85
+ - 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.
86
+ - 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.
87
+
88
+
89
+ </details>
90
+
91
+ <details>
92
+ <summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>
93
+
94
+ <!-- ### The similarity score between two dissimilar sentences is higher than 0.5 -->
95
+ **Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**
96
+
97
+ Since we finetune the models by contrastive learning with a temperature of 0.01,
98
+ the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].
99
+ So a similarity score greater than 0.5 does not indicate that the two sentences are similar.
100
+
101
+ For downstream tasks, such as passage retrieval or semantic similarity,
102
+ **what matters is the relative order of the scores, not the absolute value.**
103
+ If you need to filter similar sentences based on a similarity threshold,
104
+ please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).
105
+
106
+ </details>
107
+
108
+ <details>
109
+ <summary>3. When does the query instruction need to be used</summary>
110
+
111
+ <!-- ### When does the query instruction need to be used -->
112
+
113
+ For a retrieval task that uses short queries to find long related documents,
114
+ it is recommended to add instructions for these short queries.
115
+ **The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**
116
+ In all cases, the documents/passages do not need to add the instruction.
117
+
118
+ </details>
119
 
 
120
 
121
  ## Usage
122
 
123
+ ### Usage for Embedding Model
124
+
125
+ Here are some examples for using `bge` models with
126
  [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
127
 
128
  #### Using FlagEmbedding
 
133
 
134
  ```python
135
  from FlagEmbedding import FlagModel
136
+ sentences_1 = ["样例数据-1", "样例数据-2"]
137
+ sentences_2 = ["样例数据-3", "样例数据-4"]
138
  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:")
139
+ embeddings_1 = model.encode(sentences_1)
140
+ embeddings_2 = model.encode(sentences_2)
141
  similarity = embeddings_1 @ embeddings_2.T
142
  print(similarity)
143
 
144
+ # for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query
145
  # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
146
  queries = ['query_1', 'query_2']
147
  passages = ["样例文档-1", "样例文档-2"]
 
149
  p_embeddings = model.encode(passages)
150
  scores = q_embeddings @ p_embeddings.T
151
  ```
152
+ For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
153
 
154
+ By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.
155
+ You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.
156
 
157
 
158
  #### Using Sentence-Transformers
159
 
160
+ You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):
161
 
162
  ```
163
  pip install -U sentence-transformers
164
  ```
165
  ```python
166
  from sentence_transformers import SentenceTransformer
167
+ sentences_1 = ["样例数据-1", "样例数据-2"]
168
+ sentences_2 = ["样例数据-3", "样例数据-4"]
169
  model = SentenceTransformer('BAAI/bge-large-zh')
170
+ embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
171
+ embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
172
  similarity = embeddings_1 @ embeddings_2.T
173
  print(similarity)
174
  ```
 
195
  model_name = "BAAI/bge-small-en"
196
  model_kwargs = {'device': 'cuda'}
197
  encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
198
+ model = HuggingFaceBgeEmbeddings(
199
  model_name=model_name,
200
  model_kwargs=model_kwargs,
201
+ encode_kwargs=encode_kwargs,
202
+ query_instruction="为这个句子生成表示以用于检索相关文章:"
203
  )
204
+ model.query_instruction = "为这个句子生成表示以用于检索相关文章:"
205
  ```
206
 
207
 
208
  #### Using HuggingFace Transformers
209
 
210
+ 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.
211
 
212
  ```python
213
  from transformers import AutoTokenizer, AutoModel
 
218
  # Load model from HuggingFace Hub
219
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
220
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
221
+ model.eval()
222
 
223
  # Tokenize sentences
224
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
 
235
  print("Sentence embeddings:", sentence_embeddings)
236
  ```
237
 
238
+ ### Usage for Reranker
239
+
240
+ You can get a relevance score by inputting query and passage to the reranker.
241
+ The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.
242
+
243
+
244
+ #### Using FlagEmbedding
245
+ ```
246
+ pip install -U FlagEmbedding
247
+ ```
248
+
249
+ Get relevance score:
250
+ ```python
251
+ from FlagEmbedding import FlagReranker
252
+ reranker = FlagReranker('BAAI/bge-reranker-base', use_fp16=True) #use fp16 can speed up computing
253
+
254
+ score = reranker.compute_score(['query', 'passage'])
255
+ print(score)
256
+
257
+ 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.']])
258
+ print(scores)
259
+ ```
260
+
261
+
262
+ #### Using Huggingface transformers
263
+
264
+ ```python
265
+ import torch
266
+ from transformers import AutoModelForSequenceClassification, AutoTokenizer, BatchEncoding, PreTrainedTokenizerFast
267
+
268
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-base')
269
+ model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base')
270
+ model.eval()
271
+
272
+ 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.']]
273
+ with torch.no_grad():
274
+ inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
275
+ scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
276
+ print(scores)
277
+ ```
278
 
279
  ## Evaluation
280
+
281
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
282
+ For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
283
 
284
  - **MTEB**:
285
 
286
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
287
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
288
+ | [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 |
289
+ | [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 |
290
+ | [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 |
291
+ | [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 |
292
+ | [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 |
293
  | [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 |
294
  | [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 |
295
  | [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 |
296
+ | [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 |
297
  | [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 |
298
  | [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 |
299
  | [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 |
 
302
  | [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 |
303
  | [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 |
304
  | [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 |
 
 
 
 
305
 
306
 
307
 
308
  - **C-MTEB**:
309
+ We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
310
  Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
311
 
312
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
313
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
314
+ | [**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 |
315
+ | [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 |
316
+ | [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 |
317
+ | [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 |
318
+ | [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 |
319
+ | [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 |
320
+ | [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 |
321
+ | [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 |
322
+ | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |
323
+ | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |
324
+ | [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 |
325
+ | [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 |
326
+ | [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 |
327
+ | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |
328
+ | [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |
329
+ | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |
330
+
331
+
332
+ - **Reranking**:
333
+ See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.
334
+
335
+ | Model | T2Reranking | T2RerankingZh2En\* | T2RerankingEn2Zh\* | MmarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
336
+ |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
337
+ | text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |
338
+ | multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |
339
+ | multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |
340
+ | multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |
341
+ | m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |
342
+ | m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |
343
+ | bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |
344
+ | bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |
345
+ | [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 |
346
+ | [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 |
347
+
348
+ \* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval task
349
 
350
  ## Train
 
 
 
 
 
 
 
 
 
 
 
351
 
352
+ ### BAAI Embedding
 
 
 
 
 
 
353
 
354
+ We pre-train the models using retromae and train them on large-scale pairs data using contrastive learning.
355
+ **You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**
356
+ We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).
357
+ 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.
358
+ More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
359
 
 
 
 
 
 
 
360
 
 
 
 
361
 
362
+ ### BGE Reranker
 
 
 
 
363
 
364
+ Cross-encoder will perform full-attention over the input pair,
365
+ which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.
366
+ Therefore, it can be used to re-rank the top-k documents returned by embedding model.
367
+ We train the cross-encoder on a multilingual pair data,
368
+ The data format is the same as embedding model, so you can fine-tune it easily following our example.
369
+ More details pelease refer to [./FlagEmbedding/reranker/README.md](./FlagEmbedding/reranker/README.md)
370
 
 
371
 
372
+ ## Contact
373
+ If you have any question or suggestion related to this project, feel free to open an issue or pull request.
374
+ You also can email Shitao Xiao(stxiao@baai.ac.cn) and Zheng Liu(liuzheng@baai.ac.cn).
375
 
 
376
 
377
+ ## License
378
+ 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.
 
 
379
 
380
 
381