notifications
News
Products
For Enterprises
⇧1
Embeddings
World-class multimodal multilingual embeddings for search/RAG/agents.
Reranker
World-class neural retriever for maximizing search relevancy.
Reader
Read URLs or search the web, get better grounding for LLMs.
Segmenter
Segment long text into chunks and do tokenization.
For Power Users
PromptPerfect
Premier tool for prompt engineering
More power user tools
SceneXplain
Leading AI solution for image captions and video summaries
BestBanner
Blog to banner, without the prompts!
JinaChat
More modality, longer memory, less cost
Rationale
Ultimate AI decision-making tools
For Developers
DocArray
The data structure for multimodal data
Jina
Build multimodal AI applications on the cloud
More developer tools
CLIP-as-service
Embed images and sentences into fixed-length vectors with CLIP
Finetuner
Fine-tune embeddings on domain specific data for better search quality
JCloud
Deploy a local project as a cloud service. Radically easy, no nasty surprises.
LangChain-Serve
Langchain apps on production with Jina & FastAPI
VectorDB
A Python vector database you just need - no more, no less
DALL-E Flow
A human-in-the-Loop workflow for creating HD images from text
DiscoArt
Create compelling Disco Diffusion artworks in one line of code
ThinkGPT
Agent techniques to augment your LLM and push it beyond its limits
DevGPT
Your virtual development team
RunGPT
An open-source cloud-native of large multimodal models serving framework
Jerboa
An experimental finetuner for open-source LLMs
Company
About us
Contact sales
Intern program
Join us
open_in_new
Download logo
open_in_new
Terms & Conditions

Auto Fine-Tuning

Just tell us which domain you want your embeddings to excel in, and we automatically deliver a ready-to-use, fine-tuned embedding model for that domain.
List fine-tuned models

What is Auto Fine-Tuning?

There are three ways to specify your requirement: a general instruction, a URL, or a query-document description. Choose one.
join_right
Query-document description
Describe what the query looks like and what the matched document looks like in your domain.
public
Or, webpage URL
Refer to the content from a URL for fine-tuning.
notes
Or, general instruction
Provide a detailed description of how the fine-tuned embeddings will be used.
Select a base embedding model
Fine-tuning allows you to take a pre-trained model and adapt it to a specific task or domain by training it on a new dataset. In practice, finding effective training data is not straightforward for many users. Effective training requires more than just throwing raw PDFs, HTMLs into the model; and it is hard to get it right. Auto fine-tuning solves this problem by automatically generating effective training data using an advanced LLM agent pipeline; and fine-tuning the model within a ML workflow. You can think it as a combination of synthetic data generation and AutoML, so all you need to do is describe your target domain in natural language and let our system do the rest.

But does it work though?

Auto fine-tuning holds an auto-magical promise to deliver fine-tuned embeddings for any domain you want. But does it really work? This is a fairly reasonable doubt. We've tested it on a variety of domains and base models to find out. Check out the cherry-picked and lemon-picked results below.
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-en
Avg. improvement
arrow_upward 2%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.505 arrow_forward 0.532 arrow_upward 5%
MAP
0.352 arrow_forward 0.389 arrow_upward 10%
MRR
0.352 arrow_forward 0.389 arrow_upward 10%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from tollefj/norwegian-nli-triplets
NDCG
0.852 arrow_forward 0.867 arrow_upward 2%
MAP
0.800 arrow_forward 0.820 arrow_upward 2%
MRR
0.800 arrow_forward 0.820 arrow_upward 2%
data_usage Synthetic data generated
Total
4648
Training
4480
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-en
Avg. improvement
arrow_upward 6%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.672 arrow_forward 0.755 arrow_upward 12%
MAP
0.567 arrow_forward 0.675 arrow_upward 19%
MRR
0.567 arrow_forward 0.675 arrow_upward 19%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from mteb/askubuntudupquestions-reranking
NDCG
0.698 arrow_forward 0.722 arrow_upward 3%
MAP
0.515 arrow_forward 0.549 arrow_upward 6%
MRR
0.666 arrow_forward 0.712 arrow_upward 7%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-en
Avg. improvement
arrow_upward 9%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.727 arrow_forward 0.861 arrow_upward 18%
MAP
0.640 arrow_forward 0.814 arrow_upward 27%
MRR
0.640 arrow_forward 0.814 arrow_upward 27%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from mteb/scidocs-reranking
NDCG
0.773 arrow_forward 0.822 arrow_upward 6%
MAP
0.575 arrow_forward 0.651 arrow_upward 13%
MRR
0.823 arrow_forward 0.884 arrow_upward 7%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-zh
Avg. improvement
arrow_upward 1%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.718 arrow_forward 0.785 arrow_upward 9%
MAP
0.629 arrow_forward 0.717 arrow_upward 14%
MRR
0.629 arrow_forward 0.717 arrow_upward 14%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from C-MTEB/CMedQAv2-reranking
NDCG
0.938 arrow_forward 0.948 arrow_upward 1%
MAP
0.912 arrow_forward 0.926 arrow_upward 2%
MRR
0.920 arrow_forward 0.933 arrow_upward 1%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-en
Avg. improvement
arrow_upward 6%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.543 arrow_forward 0.579 arrow_upward 7%
MAP
0.402 arrow_forward 0.452 arrow_upward 12%
MRR
0.402 arrow_forward 0.452 arrow_upward 12%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from nc33/triplet_sbert_law2 (machine-translated to dutch)
NDCG
0.904 arrow_forward 0.948 arrow_upward 5%
MAP
0.870 arrow_forward 0.930 arrow_upward 7%
MRR
0.870 arrow_forward 0.930 arrow_upward 7%
data_usage Synthetic data generated
Total
9128
Training
8960
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-code
Avg. improvement
arrow_downward -4%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.671 arrow_forward 0.640 arrow_downward -5%
MAP
0.569 arrow_forward 0.525 arrow_downward -8%
MRR
0.569 arrow_forward 0.525 arrow_downward -8%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from mteb/stackoverflowdupquestions-reranking
NDCG
0.640 arrow_forward 0.621 arrow_downward -3%
MAP
0.530 arrow_forward 0.505 arrow_downward -5%
MRR
0.555 arrow_forward 0.532 arrow_downward -4%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-code
Avg. improvement
arrow_downward -4%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.632 arrow_forward 0.711 arrow_upward 13%
MAP
0.517 arrow_forward 0.622 arrow_upward 20%
MRR
0.517 arrow_forward 0.622 arrow_upward 20%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from mteb/stackoverflowdupquestions-reranking
NDCG
0.640 arrow_forward 0.619 arrow_downward -3%
MAP
0.530 arrow_forward 0.504 arrow_downward -5%
MRR
0.555 arrow_forward 0.525 arrow_downward -5%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model
Base model for fine-tuning
jinaai/jina-embeddings-v2-base-en
Avg. improvement
arrow_upward 1%
description Domain instruction
speed Performance on synthetic validation set before and after fine-tuning
NDCG
0.646 arrow_forward 0.729 arrow_upward 13%
MAP
0.535 arrow_forward 0.644 arrow_upward 20%
MRR
0.535 arrow_forward 0.644 arrow_upward 20%
speed Performance on held-out test set before and after fine-tuning
check
Tested on 50 random samples from mteb/askubuntudupquestions-reranking
NDCG
0.645 arrow_forward 0.650 arrow_upward 1%
MAP
0.452 arrow_forward 0.462 arrow_upward 2%
MRR
0.606 arrow_forward 0.605 arrow_downward -0%
data_usage Synthetic data generated
Total
616
Training
448
Validation
168
open_in_newDownload synthetic dataopen_in_newDownload fine-tuned model

Auto Fine-Tuning API

Get fine-tuned embeddings for any domain you want.
report_problemWe cannot generate an API key because we couldn't verify if you are human. If you believe this is an error, please contact us.
Contact
key
API Key & Billing
more_horiz
More
homehelp_outlinefaq
Describe the domain you wish to fine-tune for.
There are three ways to specify your requirement: a general instruction, a URL, or a query-document description. Choose one.
join_right
Query-document description
Describe what the query looks like and what the matched document looks like in your domain.
public
Or, webpage URL
Refer to the content from a URL for fine-tuning.
notes
Or, general instruction
Provide a detailed description of how the fine-tuned embeddings will be used.
Choose a base embedding model for fine-tuning.
Enter your API key.
Please enter the email where you want to receive the download link upon completion.
Agree to the terms and begin fine-tuning.

FAQ

Rate Limit
ProductAPI EndpointDescription Allowed RequestWithout API Key (RPM)With API Key (RPM)With Premium API Key (RPM)Average Latency (s)Token Usage Counting
Reader APIhttps://r.jina.aiConvert URL to LLM-friendly textGET/POST2020010001.6Count the number of tokens in the output response.
Reader APIhttps://s.jina.aiSearch the web and convert results to LLM-friendly textGET/POST5401007.7Count the number of tokens in the output response.
Segmenter APIhttps://segment.jina.aiTokenize and segment long textGET/POST2020010000.3Token is not counted as usage.
Embedding APIhttps://api.jina.ai/v1/embeddingsConvert text/images to fixed-length vectorsPOSTblock60keyboard_double_arrow_up300keyboard_double_arrow_up
bolt
depends on the input size
Count the number of tokens in the input request.
Reranker APIhttps://api.jina.ai/v1/rerankTokenize and segment long textPOSTblock60keyboard_double_arrow_up300keyboard_double_arrow_up
bolt
depends on the input size
Count the number of tokens in the input request.
CC BY-NC License Self-Check
Are you using our official API or official images on Azure or AWS?
done
Yes
Are you using a paid API key or free trial key?
done
Paid API key
No restrictions. Use as per your current agreement.
info
Free API key
Free trial key can be only used for non-commercial purposes. Please purchase a paid package for commercial use.
Are you using our official model images on AWS and Azure?
No restrictions. Use as per your current agreement.
close
No
Are you using these models?
jina-embeddings-v3
jina-colbert-v2
jina-reranker-v2-base-multilingual
reader-lm-0.5b/1.5b
close
No
No restrictions apply.
done
Yes
Is your use commercial?
question_mark
Not sure
Are you:
Using it for personal or hobby projects?
This is non-commercial. You can use the models freely.
A for-profit company using it internally?
This is commercial. Contact our sales team.
Contact sales
An educational institution using it for teaching?
This is typically non-commercial. You can use the models freely.
A non-profit or NGO using it for your mission?
This is typically non-commercial, but check with us if unsure.
Contact sales
Using it in a product or service you sell?
This is commercial. Contact our sales team.
Contact sales
A government entity using it for public services?
This may be commercial. Please contact us for clarification.
Contact sales
close
No
You can use the models freely.
done
Yes
Contact our sales team for licensing.
Contact sales
Auto Fine-Tuning-related common questions
How much does the Fine-tuning API cost?
The feature is currently in beta and costs 1M tokens per fine-tuned model. You can use your existing API key from the Embedding/Reranker API if it has sufficient tokens, or you can create a new API key, which includes 1M free tokens.
What do I need to input? Do I need to provide training data?
You don't need to provide any training data. Simply describe your target domain (the domain for which you want the fine-tuned embeddings to be optimized) in natural language, or use a URL as a reference, and our system will generate synthetic data to train the model.
How long does it take to fine-tune a model?
About 30 minutes.
Where are the fine-tuned models stored?
The fine-tuned models and synthetic data are stored publicly in the Hugging Face model hub.
If I provide a reference URL, how does the system use it?
The system uses the Reader API to fetch the content from the URL. It then analyzes the content to summarize the tone and domain, which it uses as guidelines for generating synthetic data. Therefore, the URL should be publicly accessible and representative of the target domain.
Can I fine-tune a model for a specific language?
Yes, you can fine-tune a model for a non-English language. The system automatically detects the language of your domain instructions and generates synthetic data accordingly. We also recommend choosing the appropriate base model for the target language. For example, if targeting a German domain, you should select the 'jina-embeddings-v2-base-de' as the base model.
Can I fine-tune non-Jina embeddings, e.g., bge-M3?
No, our fine-tuning API only supports Jina v2 models.
How do you ensure the quality of the fine-tuned models?
At the end of the fine-tuning process, the system evaluates the model using a held-out test set and reports performance metrics. You will receive an email detailing the before/after performance on this test set. You are also encouraged to evaluate the model on your own test set to ensure its quality.
How do you generate synthetic data?
The system generates synthetic data by integrating the target domain instruction you provide with LLM agents' reasoning. It produces hard negative triplets, which are essential for training high-quality embedding models. For more details, please refer to our upcoming research paper on Arxiv.
Can I keep my fine-tuned models and synthetic data private?
Currently, no. Note that this feature is still in beta. Storing the fine-tuned models and synthetic data publicly in the Hugging Face model hub helps us and the community evaluate the quality of the training. In the future, we plan to offer a private storage option.
How can I use the fine-tuned model?
Since all fine-tuned models are uploaded to Hugging Face, you can access them via SentenceTransformers by simply specifying the model name.
I never received the email with the evaluation results. What should I do?
Please check your spam folder. If you still can't find it, please contact our support team using the email address you provided.
Contact
API-related common questions
code
Can I use the same API key for embedding, reranking, reader, fine-tuning APIs?
Yes, the same API key is valid for all search foundation products from Jina AI. This includes the embedding, reranking, reader and fine-tuning APIs, with tokens shared between the all services.
code
Can I monitor the token usage of my API key?
Yes, token usage can be monitored in the 'Buy tokens' tab by entering your API key, allowing you to view the usage history and remaining tokens.
code
What should I do if I forget my API key?
If you have misplaced a topped-up key and wish to retrieve it, please contact support AT jina.ai with your registered email for assistance.
Contact
code
Do API keys expire?
No, our API keys do not have an expiration date. However, if you suspect your key has been compromised and wish to retire it or transfer its tokens to a new key, please contact our support team for assistance.
Contact
code
Why is the first request for some models slow?
This is because our serverless architecture offloads certain models during periods of low usage. The initial request activates or 'warms up' the model, which may take a few seconds. After this initial activation, subsequent requests process much more quickly.
code
Is user input data used for training your models?
We adhere to a strict privacy policy and do not use user input data for training our models.
Billing-related common questions
attach_money
Is billing based on the number of sentences or requests?
Our pricing model is based on the total number of tokens processed, allowing users the flexibility to allocate these tokens across any number of sentences, offering a cost-effective solution for diverse text analysis requirements.
attach_money
Is there a free trial available for new users?
We offer a welcoming free trial to new users, which includes one million tokens for use with any of our models, facilitated by an auto-generated API key. Once the free token limit is reached, users can easily purchase additional tokens for their API keys via the 'Buy tokens' tab.
attach_money
Are tokens charged for failed requests?
No, tokens are not deducted for failed requests.
attach_money
What payment methods are accepted?
Payments are processed through Stripe, supporting a variety of payment methods including credit cards, Google Pay, and PayPal for your convenience.
attach_money
Is invoicing available for token purchases?
Yes, an invoice will be issued to the email address associated with your Stripe account upon the purchase of tokens.
Offices
location_on
Berlin, Germany (HQ)
Prinzessinnenstraße 19-20, 10969 Berlin, Germany
Geschäftsanschrift: Leipzigerstr. 96, 10117 Berlin, Germany
location_on
Beijing, China
Level 5, Building 6, No.48 Haidian West St. Beijing Haidian, China
location_on
Shenzhen, China
402, Floor 4, Fu'an Technology Building, Shenzhen Nanshan, China
Search Foundation
Embeddings
Reranker
Reader
Segmenter
Get Jina AI API key
Rate Limit
API Status
Company
About us
Contact sales
Newsroom
Intern program
Join us
open_in_new
Download logo
open_in_new
Terms
Security
Terms & Conditions
Privacy
Manage Cookies
email
Jina AI GmbH © 2020-2024.