FlagEmbedding

Model List | FAQ | Usage | Evaluation | Train | Contact | Citation | License

More details please refer to our Github: FlagEmbedding.

English | 中文

Hiring: We're seeking experienced NLP researchers and intern students focusing on dense retrieval and retrieval-augmented LLMs. If you're interested, please feel free to reach out to us via email at zhengliu1026@gmail.com.

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

************* 🌟Updates🌟 *************

• 10/12/2023: Release LLM-Embedder, a unified embedding model to support diverse retrieval augmentation needs for LLMs. Paper :fire:
• 09/15/2023: The technical report of BGE has been released
• 09/15/2023: The massive training data of BGE has been released
• 09/12/2023: New models:
  • New reranker model: release cross-encoder models BAAI/bge-reranker-base and BAAI/bge-reranker-large, which are more powerful than embedding model. We recommend to use/fine-tune them to re-rank top-k documents returned by embedding models.
  • update embedding model: release bge-*-v1.5 embedding model to alleviate the issue of the similarity distribution, and enhance its retrieval ability without instruction.

More

• 09/07/2023: Update fine-tune code: Add script to mine hard negatives and support adding instruction during fine-tuning.
• 08/09/2023: BGE Models are integrated into Langchain, you can use it like this; C-MTEB leaderboard is available.
• 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! :tada: :tada:
• 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.

[1]: If you need to search the relevant passages in 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.

[2]: Different from the embedding model, reranker uses question and document as input and directly output similarity instead of embedding. To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models. For example, use bge embedding model to retrieve top 100 relevant documents, and then use bge reranker to re-rank the top 100 documents to get the final top-3 results.

All models have been uploaded to Huggingface Hub, and you can see them at https://huggingface.co/BAAI. If you cannot open the Huggingface Hub, you can also download the models at https://model.baai.ac.cn/models .

Frequently asked questions

1. How to fine-tune bge embedding model?

Following this example to prepare data and fine-tune your model. Some suggestions:

• Mine hard negatives following this example, which can improve the retrieval performance.
• In general, larger hyper-parameter per_device_train_batch_size brings better performance. You can expand it by enabling --fp16, --deepspeed df_config.json (df_config.json can refer to ds_config.json, --gradient_checkpointing, etc.
• 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.
• 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.

2. The similarity score between two dissimilar sentences is higher than 0.5

Suggest to use bge v1.5, which alleviates the issue of the similarity distribution.

Since we finetune the models by contrastive learning with a temperature of 0.01, 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 sentences 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).

3. When does the query instruction need to be used

For the bge-*-v1.5, we improve its retrieval ability when not using instruction. No instruction only has a slight degradation in retrieval performance compared with using instruction. So you can generate embedding without instruction in all cases for convenience.

For a retrieval task that uses short queries to find long related documents, it is recommended to add instructions for these short queries. The best method to decide whether to add instructions for queries is choosing the setting that achieves better performance on your task. In all cases, the documents/passages do not need to add the instruction.

Usage

Usage for Embedding Model

Here are some examples of using 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_1 = ["样例数据-1", "样例数据-2"]
sentences_2 = ["样例数据-3", "样例数据-4"]
model = FlagModel('BAAI/bge-large-zh-v1.5', 
                  query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章：",
                  use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation
embeddings_1 = model.encode(sentences_1)
embeddings_2 = model.encode(sentences_2)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)

# for s2p(short query to long passage) retrieval task, suggest to 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

For the value of the argument query_instruction_for_retrieval, see Model List.

By default, FlagModel will use all available GPUs when encoding. Please set os.environ["CUDA_VISIBLE_DEVICES"] to select specific GPUs. You also can set os.environ["CUDA_VISIBLE_DEVICES"]="" to make all GPUs unavailable.

Using Sentence-Transformers

You can also use the bge models with sentence-transformers:

pip install -U sentence-transformers

from sentence_transformers import SentenceTransformer
sentences_1 = ["样例数据-1", "样例数据-2"]
sentences_2 = ["样例数据-3", "样例数据-4"]
model = SentenceTransformer('BAAI/bge-large-zh-v1.5')
embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
embeddings_2 = model.encode(sentences_2, 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-v1.5')
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-large-en-v1.5"
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,
    query_instruction="为这个句子生成表示以用于检索相关文章："
)
model.query_instruction = "为这个句子生成表示以用于检索相关文章："

Using HuggingFace Transformers

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.

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-v1.5')
model = AutoModel.from_pretrained('BAAI/bge-large-zh-v1.5')
model.eval()

# 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)

Usage for Reranker

Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding. You can get a relevance score by inputting query and passage to the reranker. The reranker is optimized based cross-entropy loss, so the relevance score is not bounded to a specific range.

Using FlagEmbedding

pip install -U FlagEmbedding

Get relevance scores (higher scores indicate more relevance):

from FlagEmbedding import FlagReranker
reranker = FlagReranker('BAAI/bge-reranker-large', use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation

score = reranker.compute_score(['query', 'passage'])
print(score)

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.']])
print(scores)

Using Huggingface transformers

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-reranker-large')
model = AutoModelForSequenceClassification.from_pretrained('BAAI/bge-reranker-large')
model.eval()

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.']]
with torch.no_grad():
    inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
    scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
    print(scores)

Evaluation

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

• MTEB:

• C-MTEB:
  We create the benchmark C-MTEB for Chinese text embedding which consists of 31 datasets from 6 tasks. Please refer to C_MTEB for a detailed introduction.

• Reranking: See C_MTEB for evaluation script.

* : T2RerankingZh2En and T2RerankingEn2Zh are cross-language retrieval tasks

Train

BAAI Embedding

We pre-train the models using retromae and train them on large-scale pair data using contrastive learning. You can fine-tune the embedding model on your data following our examples. We also provide a pre-train example. 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. For more training details for bge see baai_general_embedding.

BGE Reranker

Cross-encoder will perform full-attention over the input pair, which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model. Therefore, it can be used to re-rank the top-k documents returned by embedding model. We train the cross-encoder on a multilingual pair data, The data format is the same as embedding model, so you can fine-tune it easily following our example. For more details please refer to ./FlagEmbedding/reranker/README.md

Our Contributors:

Contact

If you have any question or suggestion related to this project, feel free to open an issue or pull request. You also can email Shitao Xiao(stxiao@baai.ac.cn) and Zheng Liu(liuzheng@baai.ac.cn).

Citation

If you find this repository useful, please consider giving a star :star: and citation

@misc{bge_embedding,
      title={C-Pack: Packaged Resources To Advance General Chinese Embedding}, 
      author={Shitao Xiao and Zheng Liu and Peitian Zhang and Niklas Muennighoff},
      year={2023},
      eprint={2309.07597},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{llm_embedder,
      title={Retrieve Anything To Augment Large Language Models}, 
      author={Peitian Zhang and Shitao Xiao and Zheng Liu and Zhicheng Dou and Jian-Yun Nie},
      year={2023},
      eprint={2310.07554},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}

License

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