ldwang commited on
Commit
d7b4914
1 Parent(s): f6e1776
Files changed (1) hide show
  1. README.md +252 -103
README.md CHANGED
@@ -4,6 +4,8 @@ language:
4
  - zh
5
  pipeline_tag: sentence-similarity
6
  ---
 
 
7
  <h1 align="center">FlagEmbedding</h1>
8
 
9
 
@@ -18,81 +20,172 @@ pipeline_tag: sentence-similarity
18
  <p>
19
  </h4>
20
 
21
- More details please refer to our github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22
 
23
- [English](README.md) | [中文](README_zh.md)
24
 
25
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
26
- And it also can be used in vector database for LLMs.
27
 
28
  ************* 🌟**Updates**🌟 *************
 
 
 
 
 
29
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
30
- - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!**
31
- - 08/01/2023: We release the Chinese Massive Text Embedding Benchmark (**C-MTEB**), consisting of 31 test dataset.
32
 
33
 
34
  ## Model List
35
 
36
  `bge` is short for `BAAI general embedding`.
37
 
38
- | Model | Language | Description | query instruction for retrieval |
39
- |:-------------------------------|:--------:| :--------:| :--------:|
40
- | [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: ` |
41
- | [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: ` |
42
- | [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: ` |
43
- | [BAAI/bge-large-zh](https://huggingface.co/BAAI/bge-large-zh) | Chinese | **rank 1st** in [C-MTEB](https://github.com/FlagOpen/FlagEmbedding/tree/master/benchmark) benchmark | `为这个句子生成表示以用于检索相关文章:` |
44
- | [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/benchmark) benchmark | |
45
- | [BAAI/bge-base-zh](https://huggingface.co/BAAI/bge-base-zh) | Chinese | a base-scale model but with competitive performance | `为这个句子生成表示以用于检索相关文章:` |
46
- | [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:` |
 
 
 
 
 
 
 
 
 
 
 
 
 
47
 
48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
 
50
  ## Usage
51
 
52
- * **Using FlagEmbedding**
 
 
 
 
 
53
  ```
54
- pip install flag_embedding
55
  ```
 
 
56
  ```python
57
- from flag_embedding import FlagModel
58
- sentences = ["样例数据-1", "样例数据-2"]
 
59
  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:")
60
- embeddings = model.encode(sentences)
61
- print(embeddings)
 
 
62
 
63
- # for retrieval task, please use encode_queries() which will automatically add the instruction to each query
64
- # corpus in retrieval task can still use encode() or encode_corpus()
65
  queries = ['query_1', 'query_2']
66
- passages = ["样例段落-1", "样例段落-2"]
67
  q_embeddings = model.encode_queries(queries)
68
  p_embeddings = model.encode(passages)
69
  scores = q_embeddings @ p_embeddings.T
70
  ```
71
- The value of argument `query_instruction_for_retrieval` see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
72
 
73
- FlagModel will use all available GPUs when encoding, please set `os.environ["CUDA_VISIBLE_DEVICES"]` to choose GPU.
 
74
 
75
 
76
- * **Using Sentence-Transformers**
77
 
78
- Using this model also is easy when you have [sentence-transformers](https://www.SBERT.net) installed:
79
 
80
  ```
81
  pip install -U sentence-transformers
82
  ```
83
  ```python
84
  from sentence_transformers import SentenceTransformer
85
- sentences = ["样例数据-1", "样例数据-2"]
 
86
  model = SentenceTransformer('BAAI/bge-large-zh')
87
- embeddings = model.encode(sentences, normalize_embeddings=True)
88
- print(embeddings)
 
 
89
  ```
90
- For retrieval task,
91
- each query should start with a instruction (instructions see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list)).
 
92
  ```python
93
  from sentence_transformers import SentenceTransformer
94
- queries = ["手机开不了机怎么办?"]
95
- passages = ["样例段落-1", "样例段落-2"]
96
  instruction = "为这个句子生成表示以用于检索相关文章:"
97
 
98
  model = SentenceTransformer('BAAI/bge-large-zh')
@@ -101,9 +194,27 @@ p_embeddings = model.encode(passages, normalize_embeddings=True)
101
  scores = q_embeddings @ p_embeddings.T
102
  ```
103
 
104
- * **Using HuggingFace Transformers**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
105
 
106
- 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.
 
 
107
 
108
  ```python
109
  from transformers import AutoTokenizer, AutoModel
@@ -114,10 +225,11 @@ sentences = ["样例数据-1", "样例数据-2"]
114
  # Load model from HuggingFace Hub
115
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
116
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
 
117
 
118
  # Tokenize sentences
119
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
120
- # for retrieval task, add a instruction to query
121
  # encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')
122
 
123
  # Compute token embeddings
@@ -130,21 +242,65 @@ sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, di
130
  print("Sentence embeddings:", sentence_embeddings)
131
  ```
132
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
133
 
134
  ## Evaluation
 
135
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
136
- More details and evaluation scripts see [benchemark](benchmark/README.md).
137
 
138
  - **MTEB**:
139
 
140
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
141
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
142
- | [**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** |
143
- | [**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 |
 
 
 
144
  | [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 |
145
  | [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 |
146
  | [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 |
147
- | [**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 |
148
  | [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 |
149
  | [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 |
150
  | [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 |
@@ -153,87 +309,80 @@ More details and evaluation scripts see [benchemark](benchmark/README.md).
153
  | [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 |
154
  | [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 |
155
  | [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 |
156
- | [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 |
157
- | [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 |
158
- | [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 |
159
- | [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 |
160
 
161
 
162
 
163
  - **C-MTEB**:
164
- We create a benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
165
- Please refer to [benchemark](benchmark/README.md) for a detailed introduction.
166
 
167
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
168
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
169
- | [**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 |
170
- | [**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** |
171
- | [**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 |
172
- | [**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 |
173
- | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 |56.91 | 48.15 | 63.99 | 70.28 | 59.34 | 47.68 |
174
- | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 |54.75 | 48.64 | 64.3 | 71.22 | 59.66 | 48.88 |
175
- | [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 |
176
- | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 39.41 | 66.62 | 65.29 | 49.25 | 44.39 |
177
- | [text2vec](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 41.71 | 67.41 | 65.18 | 49.45 | 37.66 |
178
- | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 41.98 | 70.86 | 63.42 | 49.16 | 30.02 |
179
-
180
-
181
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
182
 
183
  ## Train
184
- This section will introduce the way we used to train the general embedding.
185
- The training scripts are in [flag_embedding](https://github.com/FlagOpen/FlagEmbedding/tree/master/flag_embedding/baai_general_embedding/),
186
- and we provide some examples to do [pre-train](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain/) and [fine-tune](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).
187
-
188
 
189
- **1. RetroMAE Pre-train**
190
- We pre-train the model following the method [retromae](https://github.com/staoxiao/RetroMAE),
191
- which shows promising improvement in retrieval task ([paper](https://aclanthology.org/2022.emnlp-main.35.pdf)).
192
- The pre-training was conducted on 24 A100(40G) GPUs with a batch size of 720.
193
- In retromae, the mask ratio of the encoder and decoder are 0.3, and 0.5 respectively.
194
- We used the AdamW optimizer and the learning rate is 2e-5.
195
 
196
- **Pre-training data**:
197
- - English:
198
- - [Pile](https://pile.eleuther.ai/)
199
- - [wikipedia](https://huggingface.co/datasets/wikipedia)
200
- - [msmarco](https://huggingface.co/datasets/Tevatron/msmarco-passage-corpus)
201
- - Chinese:
202
- - Subset of [wudao](https://github.com/BAAI-WuDao/Data)
203
- - [baidu-baike](https://baike.baidu.com/)
204
 
205
 
206
- **2. Finetune**
207
- We fine-tune the model using a contrastive objective.
208
- The format of input data is a triple`(query, positive, negative)`.
209
- Besides the negative in the triple, we also adopt in-batch negatives strategy.
210
- We employ the cross-device negatives sharing method to share negatives among different GPUs,
211
- which can dramatically **increase the number of negatives**.
212
 
213
- 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).
214
- We used the AdamW optimizer and the learning rate is 1e-5.
215
- The temperature for contrastive loss is 0.01.
216
 
217
- For the version with `*-instrcution`, we add instruction to the query for the retrieval task in the training.
218
- For English, the instruction is `Represent this sentence for searching relevant passages: `;
219
- For Chinese, the instruction is `为这个句子生成表示以用于检索相关文章:`.
220
- In the evaluation, the instruction should be added for sentence to passages retrieval task, not be added for other tasks.
 
 
221
 
222
 
223
- The finetune script is accessible in this repository: [flag_embedding](https://github.com/FlagOpen/FlagEmbedding/tree/master/flag_embedding/baai_general_embedding/README.md).
224
- You can easily finetune your model with it.
 
225
 
226
- **Training data**:
227
 
228
- - 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.
229
-
230
- - For Chinese, we collect 120M text pairs from [wudao](https://github.com/BAAI-WuDao/Data), [simclue](https://github.com/CLUEbenchmark/SimCLUE) and so on.
231
-
232
- **The data collection is to be released in the future.**
233
 
234
- We will continually update the embedding models and training codes,
235
- hoping to promote the development of the embedding model community.
236
 
237
 
238
- ## License
239
- FlagEmbedding is licensed under [MIT License](). The released models can be used for commercial purposes free of charge.
 
4
  - zh
5
  pipeline_tag: sentence-similarity
6
  ---
7
+
8
+
9
  <h1 align="center">FlagEmbedding</h1>
10
 
11
 
 
20
  <p>
21
  </h4>
22
 
23
+ More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
24
+
25
+
26
+ <h4 align="center">
27
+ <p>
28
+ <a href=#model-list>Model List</a> |
29
+ <a href=#frequently-asked-questions>FAQ</a> |
30
+ <a href=#usage>Usage</a> |
31
+ <a href="#evaluation">Evaluation</a> |
32
+ <a href="#train">Train</a> |
33
+ <a href="#contact">Contact</a> |
34
+ <a href="#license">License</a>
35
+ <p>
36
+ </h4>
37
+
38
 
39
+ [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
40
 
41
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
42
+ And it also can be used in vector databases for LLMs.
43
 
44
  ************* 🌟**Updates**🌟 *************
45
+ - 09/12/2023: New Release:
46
+ - **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.
47
+ - **update embedding model**: release bge-*-v1.5 embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.
48
+ - 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.
49
+ - 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).
50
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
51
+ - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:
52
+ - 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.
53
 
54
 
55
  ## Model List
56
 
57
  `bge` is short for `BAAI general embedding`.
58
 
59
+ | Model | Language | | Description | query instruction for retrieval\* |
60
+ |:-------------------------------|:--------:| :--------:| :--------:|:--------:|
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 \** | |
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 \** | |
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
+ \*: 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
+ \**: 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
 
83
+ ## Frequently asked questions
84
+
85
+ <details>
86
+ <summary>1. How to fine-tune bge embedding model?</summary>
87
+
88
+ <!-- ### How to fine-tune bge embedding model? -->
89
+ Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.
90
+ Some suggestions:
91
+ - Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#data-format), which can improve the retrieval performance.
92
+ - 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.
93
+ - 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.
94
+
95
+
96
+ </details>
97
+
98
+ <details>
99
+ <summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>
100
+
101
+ <!-- ### The similarity score between two dissimilar sentences is higher than 0.5 -->
102
+ **Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**
103
+
104
+ Since we finetune the models by contrastive learning with a temperature of 0.01,
105
+ the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].
106
+ So a similarity score greater than 0.5 does not indicate that the two sentences are similar.
107
+
108
+ For downstream tasks, such as passage retrieval or semantic similarity,
109
+ **what matters is the relative order of the scores, not the absolute value.**
110
+ If you need to filter similar sentences based on a similarity threshold,
111
+ please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).
112
+
113
+ </details>
114
+
115
+ <details>
116
+ <summary>3. When does the query instruction need to be used</summary>
117
+
118
+ <!-- ### When does the query instruction need to be used -->
119
+
120
+ For a retrieval task that uses short queries to find long related documents,
121
+ it is recommended to add instructions for these short queries.
122
+ **The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**
123
+ In all cases, the documents/passages do not need to add the instruction.
124
+
125
+ </details>
126
+
127
 
128
  ## Usage
129
 
130
+ ### Usage for Embedding Model
131
+
132
+ Here are some examples for using `bge` models with
133
+ [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
134
+
135
+ #### Using FlagEmbedding
136
  ```
137
+ pip install -U FlagEmbedding
138
  ```
139
+ 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.
140
+
141
  ```python
142
+ from FlagEmbedding import FlagModel
143
+ sentences_1 = ["样例数据-1", "样例数据-2"]
144
+ sentences_2 = ["样例数据-3", "样例数据-4"]
145
  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:")
146
+ embeddings_1 = model.encode(sentences_1)
147
+ embeddings_2 = model.encode(sentences_2)
148
+ similarity = embeddings_1 @ embeddings_2.T
149
+ print(similarity)
150
 
151
+ # for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query
152
+ # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
153
  queries = ['query_1', 'query_2']
154
+ passages = ["样例文档-1", "样例文档-2"]
155
  q_embeddings = model.encode_queries(queries)
156
  p_embeddings = model.encode(passages)
157
  scores = q_embeddings @ p_embeddings.T
158
  ```
159
+ For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
160
 
161
+ By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.
162
+ You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.
163
 
164
 
165
+ #### Using Sentence-Transformers
166
 
167
+ You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):
168
 
169
  ```
170
  pip install -U sentence-transformers
171
  ```
172
  ```python
173
  from sentence_transformers import SentenceTransformer
174
+ sentences_1 = ["样例���据-1", "样例数据-2"]
175
+ sentences_2 = ["样例数据-3", "样例数据-4"]
176
  model = SentenceTransformer('BAAI/bge-large-zh')
177
+ embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
178
+ embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
179
+ similarity = embeddings_1 @ embeddings_2.T
180
+ print(similarity)
181
  ```
182
+ For s2p(short query to long passage) retrieval task,
183
+ each short query should start with an instruction (instructions see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list)).
184
+ But the instruction is not needed for passages.
185
  ```python
186
  from sentence_transformers import SentenceTransformer
187
+ queries = ['query_1', 'query_2']
188
+ passages = ["样例文档-1", "样例文档-2"]
189
  instruction = "为这个句子生成表示以用于检索相关文章:"
190
 
191
  model = SentenceTransformer('BAAI/bge-large-zh')
 
194
  scores = q_embeddings @ p_embeddings.T
195
  ```
196
 
197
+ #### Using Langchain
198
+
199
+ You can use `bge` in langchain like this:
200
+ ```python
201
+ from langchain.embeddings import HuggingFaceBgeEmbeddings
202
+ model_name = "BAAI/bge-small-en"
203
+ model_kwargs = {'device': 'cuda'}
204
+ encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
205
+ model = HuggingFaceBgeEmbeddings(
206
+ model_name=model_name,
207
+ model_kwargs=model_kwargs,
208
+ encode_kwargs=encode_kwargs,
209
+ query_instruction="为这个句子生成表示以用于检索相关文章:"
210
+ )
211
+ model.query_instruction = "为这个句子生成表示以用于检索相关文章:"
212
+ ```
213
+
214
 
215
+ #### Using HuggingFace Transformers
216
+
217
+ 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.
218
 
219
  ```python
220
  from transformers import AutoTokenizer, AutoModel
 
225
  # Load model from HuggingFace Hub
226
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
227
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
228
+ model.eval()
229
 
230
  # Tokenize sentences
231
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
232
+ # for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)
233
  # encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')
234
 
235
  # Compute token embeddings
 
242
  print("Sentence embeddings:", sentence_embeddings)
243
  ```
244
 
245
+ ### Usage for Reranker
246
+
247
+ You can get a relevance score by inputting query and passage to the reranker.
248
+ The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.
249
+
250
+
251
+ #### Using FlagEmbedding
252
+ ```
253
+ pip install -U FlagEmbedding
254
+ ```
255
+
256
+ Get relevance score:
257
+ ```python
258
+ from FlagEmbedding import FlagReranker
259
+ reranker = FlagReranker('BAAI/bge-reranker-base', use_fp16=True) #use fp16 can speed up computing
260
+
261
+ score = reranker.compute_score(['query', 'passage'])
262
+ print(score)
263
+
264
+ 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.']])
265
+ print(scores)
266
+ ```
267
+
268
+
269
+ #### Using Huggingface transformers
270
+
271
+ ```python
272
+ import torch
273
+ from transformers import AutoModelForSequenceClassification, AutoTokenizer, BatchEncoding, PreTrainedTokenizerFast
274
+
275
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-base')
276
+ model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base')
277
+ model.eval()
278
+
279
+ 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.']]
280
+ with torch.no_grad():
281
+ inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
282
+ scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
283
+ print(scores)
284
+ ```
285
 
286
  ## Evaluation
287
+
288
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
289
+ For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
290
 
291
  - **MTEB**:
292
 
293
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
294
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
295
+ | [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 |
296
+ | [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 |
297
+ | [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 |
298
+ | [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 |
299
+ | [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 |
300
  | [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 |
301
  | [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 |
302
  | [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 |
303
+ | [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 |
304
  | [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 |
305
  | [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 |
306
  | [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 |
 
309
  | [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 |
310
  | [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 |
311
  | [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 |
 
 
 
 
312
 
313
 
314
 
315
  - **C-MTEB**:
316
+ We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
317
+ Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
318
 
319
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
320
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
321
+ | [**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 |
322
+ | [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 |
323
+ | [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 |
324
+ | [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 |
325
+ | [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 |
326
+ | [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 |
327
+ | [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 |
328
+ | [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 |
329
+ | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |
330
+ | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |
331
+ | [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 |
332
+ | [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 |
333
+ | [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 |
334
+ | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |
335
+ | [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |
336
+ | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |
337
+
338
+
339
+ - **Reranking**:
340
+ See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.
341
+
342
+ | Model | T2Reranking | T2RerankingZh2En\* | T2RerankingEn2Zh\* | MmarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
343
+ |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
344
+ | text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |
345
+ | multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |
346
+ | multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |
347
+ | multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |
348
+ | m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |
349
+ | m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |
350
+ | bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |
351
+ | bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |
352
+ | [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 |
353
+ | [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 |
354
+
355
+ \* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval task
356
 
357
  ## Train
 
 
 
 
358
 
359
+ ### BAAI Embedding
 
 
 
 
 
360
 
361
+ We pre-train the models using retromae and train them on large-scale pairs data using contrastive learning.
362
+ **You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**
363
+ We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).
364
+ 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.
365
+ More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
 
 
 
366
 
367
 
 
 
 
 
 
 
368
 
369
+ ### BGE Reranker
 
 
370
 
371
+ Cross-encoder will perform full-attention over the input pair,
372
+ which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.
373
+ Therefore, it can be used to re-rank the top-k documents returned by embedding model.
374
+ We train the cross-encoder on a multilingual pair data,
375
+ The data format is the same as embedding model, so you can fine-tune it easily following our example.
376
+ More details pelease refer to [./FlagEmbedding/reranker/README.md](./FlagEmbedding/reranker/README.md)
377
 
378
 
379
+ ## Contact
380
+ If you have any question or suggestion related to this project, feel free to open an issue or pull request.
381
+ You also can email Shitao Xiao(stxiao@baai.ac.cn) and Zheng Liu(liuzheng@baai.ac.cn).
382
 
 
383
 
384
+ ## License
385
+ 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.
 
 
 
386
 
 
 
387
 
388