E5-mistral-7b-instruct
Improving Text Embeddings with Large Language Models. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, Furu Wei, arXiv 2024
This model has 32 layers and the embedding size is 4096.
Usage
Below is an example to encode queries and passages from the MS-MARCO passage ranking dataset.
Sentence Transformers
1from sentence_transformers import SentenceTransformer
2
3model = SentenceTransformer("intfloat/e5-mistral-7b-instruct")
4# In case you want to reduce the maximum sequence length:
5model.max_seq_length = 4096
6
7queries = [
8 "how much protein should a female eat",
9 "summit define",
10]
11documents = [
12 "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
13 "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."
14]
15
16query_embeddings = model.encode(queries, prompt_name="web_search_query")
17document_embeddings = model.encode(documents)
18
19scores = (query_embeddings @ document_embeddings.T) * 100
20print(scores.tolist())
Have a look at
config_sentence_transformers.json for the prompts that are pre-configured, such as
web_search_query,
sts_query, and
summarization_query. Additionally, check out
unilm/e5/utils.py for prompts we used for evaluation. You can use these via e.g.
model.encode(queries, prompt="Instruct: Given a claim, find documents that refute the claim\nQuery: ").
Transformers
1import torch
2import torch.nn.functional as F
3
4from torch import Tensor
5from transformers import AutoTokenizer, AutoModel
6
7
8def last_token_pool(last_hidden_states: Tensor,
9 attention_mask: Tensor) -> Tensor:
10 left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
11 if left_padding:
12 return last_hidden_states[:, -1]
13 else:
14 sequence_lengths = attention_mask.sum(dim=1) - 1
15 batch_size = last_hidden_states.shape[0]
16 return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]
17
18
19def get_detailed_instruct(task_description: str, query: str) -> str:
20 return f'Instruct: {task_description}\nQuery: {query}'
21
22
23# Each query must come with a one-sentence instruction that describes the task
24task = 'Given a web search query, retrieve relevant passages that answer the query'
25queries = [
26 get_detailed_instruct(task, 'how much protein should a female eat'),
27 get_detailed_instruct(task, 'summit define')
28]
29# No need to add instruction for retrieval documents
30documents = [
31 "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
32 "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."
33]
34input_texts = queries + documents
35
36tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-mistral-7b-instruct')
37model = AutoModel.from_pretrained('intfloat/e5-mistral-7b-instruct')
38
39max_length = 4096
40# Tokenize the input texts
41batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')
42
43outputs = model(**batch_dict)
44embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
45
46# normalize embeddings
47embeddings = F.normalize(embeddings, p=2, dim=1)
48scores = (embeddings[:2] @ embeddings[2:].T) * 100
49print(scores.tolist())
Supported Languages
This model is initialized from
Mistral-7B-v0.1
and fine-tuned on a mixture of multilingual datasets.
As a result, it has some multilingual capability.
However, since Mistral-7B-v0.1 is mainly trained on English data, we recommend using this model for English only.
For multilingual use cases, please refer to
multilingual-e5-large.
MTEB Benchmark Evaluation
Check out
unilm/e5 to reproduce evaluation results
on the
BEIR and
MTEB benchmark.
FAQ
1. Do I need to add instructions to the query?
Yes, this is how the model is trained, otherwise you will see a performance degradation.
The task definition should be a one-sentence instruction that describes the task.
This is a way to customize text embeddings for different scenarios through natural language instructions.
Please check out
unilm/e5/utils.py for instructions we used for evaluation.
On the other hand, there is no need to add instructions to the document side.
2. Why are my reproduced results slightly different from reported in the model card?
Different versions of transformers and pytorch could cause negligible but non-zero performance differences.
3. Where are the LoRA-only weights?
Citation
If you find our paper or models helpful, please consider cite as follows:
1@article{wang2023improving,
2 title={Improving Text Embeddings with Large Language Models},
3 author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Yang, Linjun and Majumder, Rangan and Wei, Furu},
4 journal={arXiv preprint arXiv:2401.00368},
5 year={2023}
6}
7
8@article{wang2022text,
9 title={Text Embeddings by Weakly-Supervised Contrastive Pre-training},
10 author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
11 journal={arXiv preprint arXiv:2212.03533},
12 year={2022}
13}
Limitations
Using this model for inputs longer than 4096 tokens is not recommended.
This model's multilingual capability is still inferior to
multilingual-e5-large for some cases.
