Embedding Model

Frequently asked questions

The very poor results caused by incorrect usage

Different from other embedding models using mean pooling, BGE uses the last hidden state of [cls] as the sentence embedding: sentence_embeddings = model_output[0][:, 0]. If you use mean pooling, there will be a significant decrease in performance. Therefore, make sure to use the correct method to obtain sentence vectors. You can refer to the usage method we provide.

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 want to maintain the performance on other tasks when fine-tuning on your data, you can use LM-Cocktail to merge the fine-tuned model and the original bge model. Besides, if you want to fine-tune on multiple tasks, you also can approximate the multi-task learning via model merging as LM-Cocktail.
• 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.

Here is the way we used to fine-tune bge-large-zh-v1.5: The fine-tuning datasets consist of t2ranking, dulreader, mmarco, cmedqav2, mulit-cpr, nli-zh, ocmnli, and cmnli. For t2ranking, dulreader, and mmarco, we mine hard negatives; For nli-zh, ocmnli, and cmnli, we use the pairs whose label equal to 0 as negatives; For cmedqav2 and mulit-cpr, we randomly sample negatives. The settings of fine-tuning are: train_group_size=2, learning_rate=1e-5, max_epoch=5. We train two models: one fine-tune with --query_instruction_for_retrieval "为这个句子生成表示以用于检索相关文章：", and the other model is fine-tuned with --query_instruction_for_retrieval "", and then we merge two variants into one model to make the final model can be used both with and without instruction.

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

Using FlagEmbedding

Install:

git clone https://github.com/FlagOpen/FlagEmbedding.git
cd FlagEmbedding
pip install -e .

or:

pip install -U 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)

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

If you want to evaluate the model(or your model) on your data, you can refer to this tool.

• MTEB:

Model Name	Dimension	Sequence Length	Average (56)	Retrieval (15)	Clustering (11)	Pair Classification (3)	Reranking (4)	STS (10)	Summarization (1)	Classification (12)
BAAI/bge-large-en-v1.5	1024	512	64.23	54.29	46.08	87.12	60.03	83.11	31.61	75.97
BAAI/bge-base-en-v1.5	768	512	63.55	53.25	45.77	86.55	58.86	82.4	31.07	75.53
BAAI/bge-small-en-v1.5	384	512	62.17	51.68	43.82	84.92	58.36	81.59	30.12	74.14
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

• 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.

Model	Embedding dimension	Avg	Retrieval	STS	PairClassification	Classification	Reranking	Clustering
BAAI/bge-large-zh-v1.5	1024	64.53	70.46	56.25	81.6	69.13	65.84	48.99
BAAI/bge-base-zh-v1.5	768	63.13	69.49	53.72	79.75	68.07	65.39	47.53
BAAI/bge-small-zh-v1.5	512	57.82	61.77	49.11	70.41	63.96	60.92	44.18
BAAI/bge-large-zh	1024	64.20	71.53	54.98	78.94	68.32	65.11	48.39
bge-large-zh-noinstruct	1024	63.53	70.55	53	76.77	68.58	64.91	50.01
BAAI/bge-base-zh	768	62.96	69.53	54.12	77.5	67.07	64.91	47.63
multilingual-e5-large	1024	58.79	63.66	48.44	69.89	67.34	56.00	48.23
BAAI/bge-small-zh	512	58.27	63.07	49.45	70.35	63.64	61.48	45.09
m3e-base	768	57.10	56.91	50.47	63.99	67.52	59.34	47.68
m3e-large	1024	57.05	54.75	50.42	64.3	68.2	59.66	48.88
multilingual-e5-base	768	55.48	61.63	46.49	67.07	65.35	54.35	40.68
multilingual-e5-small	384	55.38	59.95	45.27	66.45	65.85	53.86	45.26
text-embedding-ada-002(OpenAI)	1536	53.02	52.0	43.35	69.56	64.31	54.28	45.68
luotuo	1024	49.37	44.4	42.78	66.62	61	49.25	44.39
text2vec-base	768	47.63	38.79	43.41	67.41	62.19	49.45	37.66
text2vec-large	1024	47.36	41.94	44.97	70.86	60.66	49.16	30.02

Acknowledgement

Part of the code is developed based on Dense.

Citation

If you find this repository useful, please consider giving a 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}
}