jina-embeddings-v3 / README.md
jupyterjazz's picture
readme: performance (#13)
041836f verified
|
raw
history blame
10.5 kB
metadata
license: cc-by-nc-4.0
tags:
  - feature-extraction
  - sentence-similarity
  - mteb
language:
  - multilingual
  - af
  - am
  - ar
  - as
  - az
  - be
  - bg
  - bn
  - br
  - bs
  - ca
  - cs
  - cy
  - da
  - de
  - el
  - en
  - eo
  - es
  - et
  - eu
  - fa
  - fi
  - fr
  - fy
  - ga
  - gd
  - gl
  - gu
  - ha
  - he
  - hi
  - hr
  - hu
  - hy
  - id
  - is
  - it
  - ja
  - jv
  - ka
  - kk
  - km
  - kn
  - ko
  - ku
  - ky
  - la
  - lo
  - lt
  - lv
  - mg
  - mk
  - ml
  - mn
  - mr
  - ms
  - my
  - ne
  - nl
  - 'no'
  - om
  - or
  - pa
  - pl
  - ps
  - pt
  - ro
  - ru
  - sa
  - sd
  - si
  - sk
  - sl
  - so
  - sq
  - sr
  - su
  - sv
  - sw
  - ta
  - te
  - th
  - tl
  - tr
  - ug
  - uk
  - ur
  - uz
  - vi
  - xh
  - yi
  - zh
inference: false
library_name: transformers



Finetuner logo: Finetuner helps you to create experiments in order to improve embeddings on search tasks. It accompanies you to deliver the last mile of performance-tuning for neural search applications.

The embedding set trained by Jina AI.

Jina Embedding V3: A Multilingual Multi-Task Embedding Model

Quick Start

The easiest way to start using jina-embeddings-v3 is Jina AI's Embedding API.

Intended Usage & Model Info

jina-embeddings-v3 is a multilingual multi-task text embedding model designed for a variety of NLP applications. Based on the XLM-RoBERTa architecture, this model supports Rotary Position Embeddings (RoPE) to handle long sequences up to 8192 tokens. Additionally, it features LoRA adapters to generate task-specific embeddings efficiently.

Key Features:

  • Extended Sequence Length: Supports up to 8192 tokens with RoPE.
  • Task-Specific Embedding: Customize embeddings through the task_type argument with the following options:
    • retrieval.query: Used for query embeddings in asymmetric retrieval tasks
    • retrieval.passage: Used for passage embeddings in asymmetric retrieval tasks
    • separation: Used for embeddings in clustering and re-ranking applications
    • classification: Used for embeddings in classification tasks
    • text-matching: Used for embeddings in tasks that quantify similarity between two texts, such as STS or symmetric retrieval tasks
  • Matryoshka Embeddings: Supports flexible embedding sizes (32, 64, 128, 256, 512, 768, 1024), allowing for truncating embeddings to fit your application.

Model Lineage:

jina-embeddings-v3 builds upon the FacebookAI/xlm-roberta-large model, which was originally trained on 100 languages. We extended its capabilities with an extra pretraining phase on the CulturaX dataset, then contrastively fine-tuned it on 30 languages for enhanced performance on embedding tasks in both monolingual and cross-lingual setups.

Supported Languages:

While the base model supports 100 languages, we've focused our tuning efforts on the following 30 languages: Arabic, Bengali, Chinese, Danish, Dutch, English, Finnish, French, Georgian, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Latvian, Norwegian, Polish, Portuguese, Romanian, Russian, Slovak, Spanish, Swedish, Thai, Turkish, Ukrainian, Urdu, and Vietnamese.

Data & Parameters

The data and training details are described in the technical report (coming soon).

Usage

Apply mean pooling when integrating the model.

Why Use Mean Pooling?

Mean pooling takes all token embeddings from the model's output and averages them at the sentence or paragraph level. This approach has been shown to produce high-quality sentence embeddings.

We provide an encode function that handles this for you automatically.

However, if you're working with the model directly, outside of the encode function, you'll need to apply mean pooling manually. Here's how you can do it:

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

sentences = ['How is the weather today?', 'What is the current weather like today?']

tokenizer = AutoTokenizer.from_pretrained('jinaai/jina-embeddings-v3')
model = AutoModel.from_pretrained('jinaai/jina-embeddings-v3', trust_remote_code=True)

encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

with torch.no_grad():
    model_output = model(**encoded_input)

embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
embeddings = F.normalize(embeddings, p=2, dim=1)

The easiest way to start using jina-embeddings-v3 is Jina AI's Embedding API.

Alternatively, you can use jina-embeddings-v3 directly via Transformers package:

!pip install transformers
from transformers import AutoModel

# Initialize the model
model = AutoModel.from_pretrained('jinaai/jina-embeddings-v3', trust_remote_code=True)

texts = [
    'Follow the white rabbit.',              # English
    'Sigue al conejo blanco.',               # Spanish
    'Suis le lapin blanc.',                  # French
    '跟着白兔走。',                            # Chinese
    'اتبع الأرنب الأبيض.',                     # Arabic
    'Folge dem weißen Kaninchen.'            # German
]

# When calling the `encode` function, you can choose a `task_type` based on the use case:
# 'retrieval.query', 'retrieval.passage', 'separation', 'classification', 'text-matching'
# Alternatively, you can choose not to pass a `task_type`, and no specific LoRA adapter will be used.
embeddings = model.encode(texts, task_type='text-matching')

# Compute similarities
print(embeddings[0] @ embeddings[1].T)

By default, the model supports a maximum sequence length of 8192 tokens. However, if you want to truncate your input texts to a shorter length, you can pass the max_length parameter to the encode function:

embeddings = model.encode(
    ['Very long ... document'],
    max_length=2048
)

In case you want to use Matryoshka embeddings and switch to a different dimension, you can adjust it by passing the truncate_dim parameter to the encode function:

embeddings = model.encode(
    ['Sample text'],
    truncate_dim=256
)

The latest version (#todo: specify version) of SentenceTransformers also supports jina-embeddings-v3:

!pip install -U sentence-transformers
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "jinaai/jina-embeddings-v3",
    prompts={
        "retrieval.query": "Represent the query for retrieving evidence documents: ",
        "retrieval.passage": "Represent the document for retrieval: ",
    },
    trust_remote_code=True
)

embeddings = model.encode(['What is the weather like in Berlin today?'], task_type='retrieval.query')

Performance

English MTEB

Model Average Classification Clustering Pair Classification Reranking Retrieval STS Summarization
jina-embeddings-v2-en 58.12 68.82 40.08 84.44 55.09 45.64 80.00 30.56
jina-embeddings-v3 65.60 82.58 45.27 84.01 58.13 53.87 85.8 30.98
text-embedding-3-large 62.03 75.45 49.01 84.22 59.16 55.44 81.04 29.92
multilingual-e5-large-instruct 64.41 77.56 47.1 86.19 58.58 52.47 84.78 30.39
Cohere-embed-multilingual-v3.0 60.08 64.01 46.6 86.15 57.86 53.84 83.15 30.99

Multilingual MTEB

Model Average Classification Clustering Pair Classification Reranking Retrieval STS Summarization
jina-embeddings-v2 60.54 65.69 39.36 82.95 66.57 58.24 66.6 -
jina-embeddings-v3 64.44 71.46 46.71 76.91 63.98 57.98 69.83 -
multilingual-e5-large 59.58 65.22 42.12 76.95 63.4 52.37 64.65 -
multilingual-e5-large-instruct 64.25 67.45 52.12 77.79 69.02 58.38 68.77 -

Long Context Tasks (LongEmbed)

Model Average NarrativeQA Needle Passkey QMSum SummScreen WikiQA
jina-embeddings-v3* 70.39 33.32 84.00 100.00 39.75 92.78 72.46
jina-embeddings-v2 58.12 37.89 54.25 50.25 38.87 93.48 73.99
text-embedding-3-large 51.3 44.09 29.25 63.00 32.49 84.80 54.16
baai-bge-m3 56.56 45.76 40.25 46.00 35.54 94.09 77.73

Notes:

  • *: text-matching adapter

Matryoshka Embeddings

Task 32 64 128 256 512 768 1024
Retrieval 52.54 58.54 61.64 62.72 63.16 63.30 63.35
STS 76.35 77.03 77.43 77.56 77.59 77.59 77.58

For a comprehensive evaluation and detailed metrics, please refer to the full paper available here (coming soon).

Contact

Join our Discord community and chat with other community members about ideas.

Citation

If you find jina-embeddings-v3 useful in your research, please cite the following paper: