ldwang commited on
Commit
7661484
1 Parent(s): d8d8763
Files changed (1) hide show
  1. README.md +208 -90
README.md CHANGED
@@ -2620,15 +2620,33 @@ pipeline_tag: sentence-similarity
2620
 
2621
  More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
2622
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2623
  [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
2624
 
2625
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
2626
- And it also can be used in vector database for LLMs.
2627
 
2628
  ************* 🌟**Updates**🌟 *************
2629
- - 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).
 
 
 
 
2630
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
2631
- - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!**
2632
  - 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.
2633
 
2634
 
@@ -2636,21 +2654,80 @@ And it also can be used in vector database for LLMs.
2636
 
2637
  `bge` is short for `BAAI general embedding`.
2638
 
2639
- | Model | Language | Description | query instruction for retrieval\* |
2640
- |:-------------------------------|:--------:| :--------:| :--------:|
2641
- | [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: ` |
2642
- | [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: ` |
2643
- | [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: ` |
2644
- | [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:` |
2645
- | [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 | |
2646
- | [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:` |
2647
- | [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:` |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2648
 
2649
- \*: 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.
2650
 
2651
  ## Usage
2652
 
2653
- Here are some examples to use `bge` models with
 
 
2654
  [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
2655
 
2656
  #### Using FlagEmbedding
@@ -2661,14 +2738,15 @@ If it doesn't work for you, you can see [FlagEmbedding](https://github.com/FlagO
2661
 
2662
  ```python
2663
  from FlagEmbedding import FlagModel
2664
- sentences = ["样例数据-1", "样例数据-2"]
 
2665
  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:")
2666
- embeddings_1 = model.encode(sentences)
2667
- embeddings_2 = model.encode(sentences)
2668
  similarity = embeddings_1 @ embeddings_2.T
2669
  print(similarity)
2670
 
2671
- # for s2p(short query to long passage) retrieval task, please use encode_queries() which will automatically add the instruction to each query
2672
  # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
2673
  queries = ['query_1', 'query_2']
2674
  passages = ["样例文档-1", "样例文档-2"]
@@ -2676,24 +2754,26 @@ q_embeddings = model.encode_queries(queries)
2676
  p_embeddings = model.encode(passages)
2677
  scores = q_embeddings @ p_embeddings.T
2678
  ```
2679
- The value of argument `query_instruction_for_retrieval` see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
2680
 
2681
- FlagModel will use all available GPUs when encoding, please set `os.environ["CUDA_VISIBLE_DEVICES"]` to choose GPU.
 
2682
 
2683
 
2684
  #### Using Sentence-Transformers
2685
 
2686
- Using this model also is easy when you have [sentence-transformers](https://www.SBERT.net) installed:
2687
 
2688
  ```
2689
  pip install -U sentence-transformers
2690
  ```
2691
  ```python
2692
  from sentence_transformers import SentenceTransformer
2693
- sentences = ["样例数据-1", "样例数据-2"]
 
2694
  model = SentenceTransformer('BAAI/bge-large-zh')
2695
- embeddings_1 = model.encode(sentences, normalize_embeddings=True)
2696
- embeddings_2 = model.encode(sentences, normalize_embeddings=True)
2697
  similarity = embeddings_1 @ embeddings_2.T
2698
  print(similarity)
2699
  ```
@@ -2720,17 +2800,19 @@ from langchain.embeddings import HuggingFaceBgeEmbeddings
2720
  model_name = "BAAI/bge-small-en"
2721
  model_kwargs = {'device': 'cuda'}
2722
  encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
2723
- model_norm = HuggingFaceBgeEmbeddings(
2724
  model_name=model_name,
2725
  model_kwargs=model_kwargs,
2726
- encode_kwargs=encode_kwargs
 
2727
  )
 
2728
  ```
2729
 
2730
 
2731
  #### Using HuggingFace Transformers
2732
 
2733
- 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.
2734
 
2735
  ```python
2736
  from transformers import AutoTokenizer, AutoModel
@@ -2741,6 +2823,7 @@ sentences = ["样例数据-1", "样例数据-2"]
2741
  # Load model from HuggingFace Hub
2742
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
2743
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
 
2744
 
2745
  # Tokenize sentences
2746
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
@@ -2757,21 +2840,65 @@ sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, di
2757
  print("Sentence embeddings:", sentence_embeddings)
2758
  ```
2759
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2760
 
2761
  ## Evaluation
 
2762
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
2763
- More details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
2764
 
2765
  - **MTEB**:
2766
 
2767
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
2768
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
2769
- | [**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** |
2770
- | [**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 |
 
 
 
2771
  | [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 |
2772
  | [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 |
2773
  | [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 |
2774
- | [**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 |
2775
  | [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 |
2776
  | [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 |
2777
  | [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 |
@@ -2780,89 +2907,80 @@ More details and evaluation tools see our [scripts](https://github.com/FlagOpen/
2780
  | [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 |
2781
  | [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 |
2782
  | [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 |
2783
- | [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 |
2784
- | [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 |
2785
- | [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 |
2786
- | [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 |
2787
 
2788
 
2789
 
2790
  - **C-MTEB**:
2791
- We create a benchmark C-MTEB for chinese text embedding which consists of 31 datasets from 6 tasks.
2792
  Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
2793
 
2794
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
2795
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
2796
- | [**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 |
2797
- | [**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** |
2798
- | [**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 |
2799
- | [**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 |
2800
- | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 |56.91 | 48.15 | 63.99 | 70.28 | 59.34 | 47.68 |
2801
- | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 |54.75 | 48.64 | 64.3 | 71.22 | 59.66 | 48.88 |
2802
- | [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 |
2803
- | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 39.41 | 66.62 | 65.29 | 49.25 | 44.39 |
2804
- | [text2vec](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 41.71 | 67.41 | 65.18 | 49.45 | 37.66 |
2805
- | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 41.98 | 70.86 | 63.42 | 49.16 | 30.02 |
2806
-
 
 
 
 
 
2807
 
2808
 
2809
- ## Train
2810
- This section will introduce the way we used to train the general embedding.
2811
- The training scripts are in [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md),
2812
- 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).
2813
-
2814
 
2815
- **1. RetroMAE Pre-train**
2816
- We pre-train the model following the method [retromae](https://github.com/staoxiao/RetroMAE),
2817
- which shows promising improvement in retrieval task ([paper](https://aclanthology.org/2022.emnlp-main.35.pdf)).
2818
- The pre-training was conducted on 24 A100(40G) GPUs with a batch size of 720.
2819
- In retromae, the mask ratio of encoder and decoder are 0.3, 0.5 respectively.
2820
- We used the AdamW optimizer and the learning rate is 2e-5.
 
 
 
 
 
 
2821
 
2822
- **Pre-training data**:
2823
- - English:
2824
- - [Pile](https://pile.eleuther.ai/)
2825
- - [wikipedia](https://huggingface.co/datasets/wikipedia)
2826
- - [msmarco](https://huggingface.co/datasets/Tevatron/msmarco-passage-corpus)
2827
- - Chinese:
2828
- - [wudao](https://github.com/BAAI-WuDao/Data)
2829
 
 
2830
 
2831
- **2. Finetune**
2832
- We fine-tune the model using a contrastive objective.
2833
- The format of input data is a triple`(query, positive, negative)`.
2834
- Besides the negative in the triple, we also adopt in-batch negatives strategy.
2835
- We employ the cross-device negatives sharing method to share negatives among different GPUs,
2836
- which can dramatically **increase the number of negatives**.
2837
-
2838
- 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).
2839
- We used the AdamW optimizer and the learning rate is 1e-5.
2840
- The temperature for contrastive loss is 0.01.
2841
-
2842
- Besides, we add instruction to the query for s2p(short query to long passage) retrieval task in the training (add nothing to passages).
2843
- For English, the instruction is `Represent this sentence for searching relevant passages: `;
2844
- For Chinese, the instruction is `为这个句子生成表示以用于检索相关文章:`.
2845
- In the evaluation, the instruction should be added for queries in retrieval task, not be added for other tasks.
2846
- Noted that the instruction is not needed for passages.
2847
 
2848
- The finetune script is accessible in this repository: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
2849
- You can easily finetune your model with it.
 
 
 
2850
 
2851
- **Training data**:
2852
 
2853
- - 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.
2854
 
2855
- - For chinese, we collect 120M text pairs from [wudao](https://github.com/BAAI-WuDao/Data), [simclue](https://github.com/CLUEbenchmark/SimCLUE) and so on.
2856
 
2857
- **The data collection is to be released in the future.**
 
 
 
 
 
2858
 
2859
- We will continually update the embedding models and training codes,
2860
- hoping to promote the development of the embedding model community.
2861
 
 
 
 
2862
 
2863
 
2864
  ## License
2865
- 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.
2866
 
2867
 
2868
 
 
2620
 
2621
  More details please refer to our Github: [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding).
2622
 
2623
+
2624
+ <h4 align="center">
2625
+ <p>
2626
+ <a href=#model-list>Model List</a> |
2627
+ <a href=#frequently-asked-questions>FAQ</a> |
2628
+ <a href=#usage>Usage</a> |
2629
+ <a href="#evaluation">Evaluation</a> |
2630
+ <a href="#train">Train</a> |
2631
+ <a href="#contact">Contact</a> |
2632
+ <a href="#license">License</a>
2633
+ <p>
2634
+ </h4>
2635
+
2636
+
2637
  [English](README.md) | [中文](https://github.com/FlagOpen/FlagEmbedding/blob/master/README_zh.md)
2638
 
2639
  FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search.
2640
+ And it also can be used in vector databases for LLMs.
2641
 
2642
  ************* 🌟**Updates**🌟 *************
2643
+ - 09/12/2023: New Release:
2644
+ - **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.
2645
+ - **update embedding model**: release bge-*-v1.5 embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.
2646
+ - 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.
2647
+ - 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).
2648
  - 08/05/2023: Release base-scale and small-scale models, **best performance among the models of the same size 🤗**
2649
+ - 08/02/2023: Release `bge-large-*`(short for BAAI General Embedding) Models, **rank 1st on MTEB and C-MTEB benchmark!** :tada: :tada:
2650
  - 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.
2651
 
2652
 
 
2654
 
2655
  `bge` is short for `BAAI general embedding`.
2656
 
2657
+ | Model | Language | | Description | query instruction for retrieval\* |
2658
+ |:-------------------------------|:--------:| :--------:| :--------:|:--------:|
2659
+ | [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 \** | |
2660
+ | [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 \** | |
2661
+ | [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: ` |
2662
+ | [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: ` |
2663
+ | [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: ` |
2664
+ | [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:` |
2665
+ | [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:` |
2666
+ | [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:` |
2667
+ | [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: ` |
2668
+ | [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: ` |
2669
+ | [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: ` |
2670
+ | [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:` |
2671
+ | [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:` |
2672
+ | [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:` |
2673
+
2674
+
2675
+ \*: 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.
2676
+
2677
+ \**: To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models.
2678
+ 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.
2679
+
2680
+
2681
+ ## Frequently asked questions
2682
+
2683
+ <details>
2684
+ <summary>1. How to fine-tune bge embedding model?</summary>
2685
+
2686
+ <!-- ### How to fine-tune bge embedding model? -->
2687
+ Following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune) to prepare data and fine-tune your model.
2688
+ Some suggestions:
2689
+ - Mine hard negatives following this [example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune#data-format), which can improve the retrieval performance.
2690
+ - 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.
2691
+ - 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.
2692
+
2693
+
2694
+ </details>
2695
+
2696
+ <details>
2697
+ <summary>2. The similarity score between two dissimilar sentences is higher than 0.5</summary>
2698
+
2699
+ <!-- ### The similarity score between two dissimilar sentences is higher than 0.5 -->
2700
+ **Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.**
2701
+
2702
+ Since we finetune the models by contrastive learning with a temperature of 0.01,
2703
+ the similarity distribution of the current BGE model is about in the interval \[0.6, 1\].
2704
+ So a similarity score greater than 0.5 does not indicate that the two sentences are similar.
2705
+
2706
+ For downstream tasks, such as passage retrieval or semantic similarity,
2707
+ **what matters is the relative order of the scores, not the absolute value.**
2708
+ If you need to filter similar sentences based on a similarity threshold,
2709
+ please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).
2710
+
2711
+ </details>
2712
+
2713
+ <details>
2714
+ <summary>3. When does the query instruction need to be used</summary>
2715
+
2716
+ <!-- ### When does the query instruction need to be used -->
2717
+
2718
+ For a retrieval task that uses short queries to find long related documents,
2719
+ it is recommended to add instructions for these short queries.
2720
+ **The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task.**
2721
+ In all cases, the documents/passages do not need to add the instruction.
2722
+
2723
+ </details>
2724
 
 
2725
 
2726
  ## Usage
2727
 
2728
+ ### Usage for Embedding Model
2729
+
2730
+ Here are some examples for using `bge` models with
2731
  [FlagEmbedding](#using-flagembedding), [Sentence-Transformers](#using-sentence-transformers), [Langchain](#using-langchain), or [Huggingface Transformers](#using-huggingface-transformers).
2732
 
2733
  #### Using FlagEmbedding
 
2738
 
2739
  ```python
2740
  from FlagEmbedding import FlagModel
2741
+ sentences_1 = ["样例数据-1", "样例数据-2"]
2742
+ sentences_2 = ["样例数据-3", "样例数据-4"]
2743
  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:")
2744
+ embeddings_1 = model.encode(sentences_1)
2745
+ embeddings_2 = model.encode(sentences_2)
2746
  similarity = embeddings_1 @ embeddings_2.T
2747
  print(similarity)
2748
 
2749
+ # for s2p(short query to long passage) retrieval task, suggest to use encode_queries() which will automatically add the instruction to each query
2750
  # corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
2751
  queries = ['query_1', 'query_2']
2752
  passages = ["样例文档-1", "样例文档-2"]
 
2754
  p_embeddings = model.encode(passages)
2755
  scores = q_embeddings @ p_embeddings.T
2756
  ```
2757
+ For the value of the argument `query_instruction_for_retrieval`, see [Model List](https://github.com/FlagOpen/FlagEmbedding/tree/master#model-list).
2758
 
2759
+ By default, FlagModel will use all available GPUs when encoding. Please set `os.environ["CUDA_VISIBLE_DEVICES"]` to select specific GPUs.
2760
+ You also can set `os.environ["CUDA_VISIBLE_DEVICES"]=""` to make all GPUs unavailable.
2761
 
2762
 
2763
  #### Using Sentence-Transformers
2764
 
2765
+ You can also use the `bge` models with [sentence-transformers](https://www.SBERT.net):
2766
 
2767
  ```
2768
  pip install -U sentence-transformers
2769
  ```
2770
  ```python
2771
  from sentence_transformers import SentenceTransformer
2772
+ sentences_1 = ["样例数据-1", "样例数据-2"]
2773
+ sentences_2 = ["样例数据-3", "样例数据-4"]
2774
  model = SentenceTransformer('BAAI/bge-large-zh')
2775
+ embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
2776
+ embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
2777
  similarity = embeddings_1 @ embeddings_2.T
2778
  print(similarity)
2779
  ```
 
2800
  model_name = "BAAI/bge-small-en"
2801
  model_kwargs = {'device': 'cuda'}
2802
  encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
2803
+ model = HuggingFaceBgeEmbeddings(
2804
  model_name=model_name,
2805
  model_kwargs=model_kwargs,
2806
+ encode_kwargs=encode_kwargs,
2807
+ query_instruction="为这个句子生成表示以用于检索相关文章:"
2808
  )
2809
+ model.query_instruction = "为这个句子生成表示以用于检索相关文章:"
2810
  ```
2811
 
2812
 
2813
  #### Using HuggingFace Transformers
2814
 
2815
+ 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.
2816
 
2817
  ```python
2818
  from transformers import AutoTokenizer, AutoModel
 
2823
  # Load model from HuggingFace Hub
2824
  tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
2825
  model = AutoModel.from_pretrained('BAAI/bge-large-zh')
2826
+ model.eval()
2827
 
2828
  # Tokenize sentences
2829
  encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
 
2840
  print("Sentence embeddings:", sentence_embeddings)
2841
  ```
2842
 
2843
+ ### Usage for Reranker
2844
+
2845
+ You can get a relevance score by inputting query and passage to the reranker.
2846
+ The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.
2847
+
2848
+
2849
+ #### Using FlagEmbedding
2850
+ ```
2851
+ pip install -U FlagEmbedding
2852
+ ```
2853
+
2854
+ Get relevance score:
2855
+ ```python
2856
+ from FlagEmbedding import FlagReranker
2857
+ reranker = FlagReranker('BAAI/bge-reranker-base', use_fp16=True) #use fp16 can speed up computing
2858
+
2859
+ score = reranker.compute_score(['query', 'passage'])
2860
+ print(score)
2861
+
2862
+ 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.']])
2863
+ print(scores)
2864
+ ```
2865
+
2866
+
2867
+ #### Using Huggingface transformers
2868
+
2869
+ ```python
2870
+ import torch
2871
+ from transformers import AutoModelForSequenceClassification, AutoTokenizer, BatchEncoding, PreTrainedTokenizerFast
2872
+
2873
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-base')
2874
+ model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-base')
2875
+ model.eval()
2876
+
2877
+ 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.']]
2878
+ with torch.no_grad():
2879
+ inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
2880
+ scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
2881
+ print(scores)
2882
+ ```
2883
 
2884
  ## Evaluation
2885
+
2886
  `baai-general-embedding` models achieve **state-of-the-art performance on both MTEB and C-MTEB leaderboard!**
2887
+ For more details and evaluation tools see our [scripts](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md).
2888
 
2889
  - **MTEB**:
2890
 
2891
  | Model Name | Dimension | Sequence Length | Average (56) | Retrieval (15) |Clustering (11) | Pair Classification (3) | Reranking (4) | STS (10) | Summarization (1) | Classification (12) |
2892
  |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
2893
+ | [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 |
2894
+ | [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 |
2895
+ | [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 |
2896
+ | [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 |
2897
+ | [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 |
2898
  | [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 |
2899
  | [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 |
2900
  | [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 |
2901
+ | [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 |
2902
  | [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 |
2903
  | [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 |
2904
  | [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 |
 
2907
  | [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 |
2908
  | [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 |
2909
  | [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 |
 
 
 
 
2910
 
2911
 
2912
 
2913
  - **C-MTEB**:
2914
+ We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks.
2915
  Please refer to [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/README.md) for a detailed introduction.
2916
 
2917
  | Model | Embedding dimension | Avg | Retrieval | STS | PairClassification | Classification | Reranking | Clustering |
2918
  |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
2919
+ | [**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 |
2920
+ | [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 |
2921
+ | [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 |
2922
+ | [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 |
2923
+ | [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 |
2924
+ | [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 |
2925
+ | [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 |
2926
+ | [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 |
2927
+ | [m3e-base](https://huggingface.co/moka-ai/m3e-base) | 768 | 57.10 | 56.91 | 50.47 | 63.99 | 67.52 | 59.34 | 47.68 |
2928
+ | [m3e-large](https://huggingface.co/moka-ai/m3e-large) | 1024 | 57.05 | 54.75 | 50.42 | 64.3 | 68.2 | 59.66 | 48.88 |
2929
+ | [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 |
2930
+ | [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 |
2931
+ | [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 |
2932
+ | [luotuo](https://huggingface.co/silk-road/luotuo-bert-medium) | 1024 | 49.37 | 44.4 | 42.78 | 66.62 | 61 | 49.25 | 44.39 |
2933
+ | [text2vec-base](https://huggingface.co/shibing624/text2vec-base-chinese) | 768 | 47.63 | 38.79 | 43.41 | 67.41 | 62.19 | 49.45 | 37.66 |
2934
+ | [text2vec-large](https://huggingface.co/GanymedeNil/text2vec-large-chinese) | 1024 | 47.36 | 41.94 | 44.97 | 70.86 | 60.66 | 49.16 | 30.02 |
2935
 
2936
 
2937
+ - **Reranking**:
2938
+ See [C_MTEB](https://github.com/FlagOpen/FlagEmbedding/blob/master/C_MTEB/) for evaluation script.
 
 
 
2939
 
2940
+ | Model | T2Reranking | T2RerankingZh2En\* | T2RerankingEn2Zh\* | MmarcoReranking | CMedQAv1 | CMedQAv2 | Avg |
2941
+ |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|
2942
+ | text2vec-base-multilingual | 64.66 | 62.94 | 62.51 | 14.37 | 48.46 | 48.6 | 50.26 |
2943
+ | multilingual-e5-small | 65.62 | 60.94 | 56.41 | 29.91 | 67.26 | 66.54 | 57.78 |
2944
+ | multilingual-e5-large | 64.55 | 61.61 | 54.28 | 28.6 | 67.42 | 67.92 | 57.4 |
2945
+ | multilingual-e5-base | 64.21 | 62.13 | 54.68 | 29.5 | 66.23 | 66.98 | 57.29 |
2946
+ | m3e-base | 66.03 | 62.74 | 56.07 | 17.51 | 77.05 | 76.76 | 59.36 |
2947
+ | m3e-large | 66.13 | 62.72 | 56.1 | 16.46 | 77.76 | 78.27 | 59.57 |
2948
+ | bge-base-zh-v1.5 | 66.49 | 63.25 | 57.02 | 29.74 | 80.47 | 84.88 | 63.64 |
2949
+ | bge-large-zh-v1.5 | 65.74 | 63.39 | 57.03 | 28.74 | 83.45 | 85.44 | 63.97 |
2950
+ | [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 |
2951
+ | [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 |
2952
 
2953
+ \* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval task
 
 
 
 
 
 
2954
 
2955
+ ## Train
2956
 
2957
+ ### BAAI Embedding
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2958
 
2959
+ We pre-train the models using retromae and train them on large-scale pairs data using contrastive learning.
2960
+ **You can fine-tune the embedding model on your data following our [examples](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/finetune).**
2961
+ We also provide a [pre-train example](https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain).
2962
+ 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.
2963
+ More training details for bge see [baai_general_embedding](https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/README.md).
2964
 
 
2965
 
 
2966
 
2967
+ ### BGE Reranker
2968
 
2969
+ Cross-encoder will perform full-attention over the input pair,
2970
+ which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model.
2971
+ Therefore, it can be used to re-rank the top-k documents returned by embedding model.
2972
+ We train the cross-encoder on a multilingual pair data,
2973
+ The data format is the same as embedding model, so you can fine-tune it easily following our example.
2974
+ More details pelease refer to [./FlagEmbedding/reranker/README.md](./FlagEmbedding/reranker/README.md)
2975
 
 
 
2976
 
2977
+ ## Contact
2978
+ If you have any question or suggestion related to this project, feel free to open an issue or pull request.
2979
+ You also can email Shitao Xiao(stxiao@baai.ac.cn) and Zheng Liu(liuzheng@baai.ac.cn).
2980
 
2981
 
2982
  ## License
2983
+ 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.
2984
 
2985
 
2986