Git Product home page Git Product logo

flagembedding's Introduction

FlagEmbedding

Build License Build Build

English | 中文

FlagEmbedding can map any text to a low-dimensional dense vector which can be used for tasks like retrieval, classification, clustering, or semantic search. And it also can be used in vector database for LLMs.

************* 🌟Updates🌟 *************

  • 08/09/2023: BGE Models are integrated into Langchain, you can use it like this; C-MTEB leaderboard is avaliable.
  • 08/05/2023: Release base-scale and small-scale models, best performance among the models of the same size 🤗
  • 08/02/2023: Release bge-large-*(short for BAAI General Embedding) Models, rank 1st on MTEB and C-MTEB benchmark! 🎉 🎉
  • 08/01/2023: We release the Chinese Massive Text Embedding Benchmark (C-MTEB), consisting of 31 test dataset.

Model List

bge is short for BAAI general embedding.

Model Language Description query instruction for retrieval*
BAAI/bge-large-en English 🏆 rank 1st in MTEB leaderboard Represent this sentence for searching relevant passages:
BAAI/bge-base-en English rank 2nd in MTEB leaderboard Represent this sentence for searching relevant passages:
BAAI/bge-small-en English a small-scale model but with competitive performance Represent this sentence for searching relevant passages:
BAAI/bge-large-zh Chinese 🏆 rank 1st in C-MTEB benchmark 为这个句子生成表示以用于检索相关文章:
BAAI/bge-large-zh-noinstruct Chinese This model is trained without instruction, and rank 2nd in C-MTEB benchmark
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:
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:

*: 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.

Frequently asked questions

  1. The similarity score between two dissimilar sentence is higher than 0.5

The similarity distribution of the current BGE model is about in the interval [0.6, 1]. So a similarity score greater than 0.5 does not indicate that the two sentence are similar.

For downstream tasks, such as passage retrieval or semantic similarity, what matters is the relative order of the scores, not the absolute value. If you need to filter similar sentences based on a similarity threshold, please select an appropriate similarity threshold based on the similarity distribution on your data (such as 0.8, 0.85, or even 0.9).

  1. When do the query instruction need to be used

For a retrieval task that uses short queries to find long related documents, it is recommended to add instructions for these short queries. For other tasks, it is recommended not to add instructions. For example, in Quora task, which needs to use a short question to search another related short questions, the instruction is not recommended to add. The best method to decide whether to add instructions for queries is choosing the setting which can achieve better performance in your task. In all cases, the documents/passages do not need to add the instruction, only need to consider whether to add the instruction for queries.

Usage

Here are some examples to use bge models with FlagEmbedding, Sentence-Transformers, Langchain, or Huggingface Transformers.

Using FlagEmbedding

pip install -U FlagEmbedding

If it doesn't work for you, you can see FlagEmbedding for more methods to install FlagEmbedding.

from FlagEmbedding import FlagModel
sentences = ["样例数据-1", "样例数据-2"]
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:")
embeddings_1 = model.encode(sentences)
embeddings_2 = model.encode(sentences)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)

# for s2p(short query to long passage) retrieval task, please use encode_queries() which will automatically add the instruction to each query
# corpus in retrieval task can still use encode() or encode_corpus(), since they don't need instruction
queries = ['query_1', 'query_2']
passages = ["样例文档-1", "样例文档-2"]
q_embeddings = model.encode_queries(queries)
p_embeddings = model.encode(passages)
scores = q_embeddings @ p_embeddings.T

The value of argument query_instruction_for_retrieval see Model List.

FlagModel will use all available GPUs when encoding, please set os.environ["CUDA_VISIBLE_DEVICES"] to choose GPU. You also can set os.environ["CUDA_VISIBLE_DEVICES"]="" to make GPUs unavailable.

Using Sentence-Transformers

Using this model also is easy when you have sentence-transformers installed:

pip install -U sentence-transformers
from sentence_transformers import SentenceTransformer
sentences = ["样例数据-1", "样例数据-2"]
model = SentenceTransformer('BAAI/bge-large-zh')
embeddings_1 = model.encode(sentences, normalize_embeddings=True)
embeddings_2 = model.encode(sentences, normalize_embeddings=True)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)

For s2p(short query to long passage) retrieval task, each short query should start with an instruction (instructions see Model List). But the instruction is not needed for passages.

from sentence_transformers import SentenceTransformer
queries = ['query_1', 'query_2']
passages = ["样例文档-1", "样例文档-2"]
instruction = "为这个句子生成表示以用于检索相关文章:"

model = SentenceTransformer('BAAI/bge-large-zh')
q_embeddings = model.encode([instruction+q for q in queries], normalize_embeddings=True)
p_embeddings = model.encode(passages, normalize_embeddings=True)
scores = q_embeddings @ p_embeddings.T

Using Langchain

You can use bge in langchain like this:

from langchain.embeddings import HuggingFaceBgeEmbeddings
model_name = "BAAI/bge-small-en"
model_kwargs = {'device': 'cuda'}
encode_kwargs = {'normalize_embeddings': True} # set True to compute cosine similarity
model = HuggingFaceBgeEmbeddings(
    model_name=model_name,
    model_kwargs=model_kwargs,
    encode_kwargs=encode_kwargs
)

Using HuggingFace Transformers

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.

from transformers import AutoTokenizer, AutoModel
import torch
# Sentences we want sentence embeddings for
sentences = ["样例数据-1", "样例数据-2"]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh')
model = AutoModel.from_pretrained('BAAI/bge-large-zh')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
# for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)
# encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)
    # Perform pooling. In this case, cls pooling.
    sentence_embeddings = model_output[0][:, 0]
# normalize embeddings
sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
print("Sentence embeddings:", sentence_embeddings)

Evaluation

baai-general-embedding models achieve state-of-the-art performance on both MTEB and C-MTEB leaderboard! More details and evaluation tools see our scripts.

  • MTEB:
Model Name Dimension Sequence Length Average (56) Retrieval (15) Clustering (11) Pair Classification (3) Reranking (4) STS (10) Summarization (1) Classification (12)
bge-large-en 1024 512 63.98 53.9 46.98 85.8 59.48 81.56 32.06 76.21
bge-base-en 768 512 63.36 53.0 46.32 85.86 58.7 81.84 29.27 75.27
gte-large 1024 512 63.13 52.22 46.84 85.00 59.13 83.35 31.66 73.33
gte-base 768 512 62.39 51.14 46.2 84.57 58.61 82.3 31.17 73.01
e5-large-v2 1024 512 62.25 50.56 44.49 86.03 56.61 82.05 30.19 75.24
bge-small-en 384 512 62.11 51.82 44.31 83.78 57.97 80.72 30.53 74.37
instructor-xl 768 512 61.79 49.26 44.74 86.62 57.29 83.06 32.32 61.79
e5-base-v2 768 512 61.5 50.29 43.80 85.73 55.91 81.05 30.28 73.84
gte-small 384 512 61.36 49.46 44.89 83.54 57.7 82.07 30.42 72.31
text-embedding-ada-002 1536 8192 60.99 49.25 45.9 84.89 56.32 80.97 30.8 70.93
e5-small-v2 384 512 59.93 49.04 39.92 84.67 54.32 80.39 31.16 72.94
sentence-t5-xxl 768 512 59.51 42.24 43.72 85.06 56.42 82.63 30.08 73.42
all-mpnet-base-v2 768 514 57.78 43.81 43.69 83.04 59.36 80.28 27.49 65.07
sgpt-bloom-7b1-msmarco 4096 2048 57.59 48.22 38.93 81.9 55.65 77.74 33.6 66.19
all-MiniLM-L12-v2 384 512 56.53 42.69 41.81 82.41 58.44 79.8 27.9 63.21
all-MiniLM-L6-v2 384 512 56.26 41.95 42.35 82.37 58.04 78.9 30.81 63.05
contriever-base-msmarco 768 512 56.00 41.88 41.1 82.54 53.14 76.51 30.36 66.68
sentence-t5-base 768 512 55.27 33.63 40.21 85.18 53.09 81.14 31.39 69.81
  • C-MTEB:
    We create a benchmark C-MTEB for chinese text embedding which consists of 31 datasets from 6 tasks. Please refer to C_MTEB for a detailed introduction.
Model Embedding dimension Avg Retrieval STS PairClassification Classification Reranking Clustering
bge-large-zh 1024 64.20 71.53 53.23 78.94 72.26 65.11 48.39
bge-large-zh-noinstruct 1024 63.53 70.55 50.98 76.77 72.49 64.91 50.01
BAAI/bge-base-zh 768 62.96 69.53 52.05 77.5 70.98 64.91 47.63
BAAI/bge-small-zh 512 58.27 63.07 46.87 70.35 67.78 61.48 45.09
m3e-base 768 57.10 56.91 48.15 63.99 70.28 59.34 47.68
m3e-large 1024 57.05 54.75 48.64 64.3 71.22 59.66 48.88
text-embedding-ada-002(OpenAI) 1536 53.02 52.0 40.61 69.56 67.38 54.28 45.68
luotuo 1024 49.37 44.4 39.41 66.62 65.29 49.25 44.39
text2vec 768 47.63 38.79 41.71 67.41 65.18 49.45 37.66
text2vec-large 1024 47.36 41.94 41.98 70.86 63.42 49.16 30.02

Train

This section will introduce the way we used to train the general embedding. The training scripts are in FlagEmbedding, and we provide some examples to do pre-train and fine-tune.

1. RetroMAE Pre-train
We pre-train the model following the method retromae, which shows promising improvement in retrieval task (paper). The pre-training was conducted on 24 A100(40G) GPUs with a batch size of 720. In retromae, the mask ratio of encoder and decoder are 0.3, 0.5 respectively. We used the AdamW optimizer and the learning rate is 2e-5.

Pre-training data:

2. Finetune
We fine-tune the model using a contrastive objective. The format of input data is a triple(query, positive, negative). Besides the negative in the triple, we also adopt in-batch negatives strategy. We employ the cross-device negatives sharing method to share negatives among different GPUs, which can dramatically increase the number of negatives.

We trained our model on 48 A100(40G) GPUs with a large batch size of 32,784 (so there are 65,567 negatives for each query in a batch). We used the AdamW optimizer and the learning rate is 1e-5. The temperature for contrastive loss is 0.01.

Besides, we add instruction to the query for s2p(short query to long passage) retrieval task in the training (add nothing to passages). For English, the instruction is Represent this sentence for searching relevant passages: ; For Chinese, the instruction is 为这个句子生成表示以用于检索相关文章:. In the evaluation, the instruction should be added for queries in retrieval task, not be added for other tasks. Noted that the instruction is not needed for passages.

The finetune script is accessible in this repository: FlagEmbedding. You can easily finetune your model with it.

Training data:

  • For English, we collect 230M text pairs from wikipedia, cc-net, and so on.

  • For chinese, we collect 120M text pairs from wudao, simclue, and so on.

The data collection is to be released in the future.

Schedule

  • Chinese Massive Text Embedding Benchmark
  • release baai-general-embedding models
  • release codes for training
  • Multilingual model
  • Training Datasets
  • ...

We will continually update the embedding models and training codes, hoping to promote the development of the embedding model community.

Contact

If you have any question or suggestion related to this project, feel free to open an issue or pull a request. You also can email Shitao Xiao([email protected]) and Zheng Liu([email protected]).

License

FlagEmbedding is licensed under MIT License. The released models can be used for commercial purposes free of charge.

flagembedding's People

Contributors

ng-fukgin avatar staoxiao avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.