Data Synthesis#

Open In Colab

When using Finetuner, each item in your training data must either have a label, or have a similarity score comparing it to some other item. See the Finetuner documentation on preparing training data. If your data is not labelled, and you don’t want to spend time manually organizing and labelling it, you can use the finetuner.synthesize function to automatically construct a dataset that can be used in training.

This guide will walk you through the process of using the finetuner.synthesize function, as well as how to use its output for training.



!pip install 'finetuner[full]'

Prepare Synthesis Data#

To perform synthesis, we need a query dataset and a corpus dataset, with the query dataset containing examples of user queries, and the corpus containing example search results.

We’ll be generating training data based on the electronics section of the Amazon cross-market dataset, a collection of products, ratings and reviews taken from Amazon. For our purposes, we will only be using the product names.

We use the xmarket_queries_da and xmarket_corpus_da datasets, which we have already pre-processed and made available on the Jina AI Cloud. You can access them using DocumentArray.pull:

import finetuner
from docarray import Document, DocumentArray

query_data = DocumentArray.pull('finetuner/xmarket_queries_da')
corpus_data = DocumentArray.pull('finetuner/xmarket_corpus_da')


The format of the data in these DocumentArrays is very simple, each Document wraps a single item, contained in its text field.

Choosing models#

Data synthesis jobs require two different models: a relation miner and a cross encoder.

The relation miner is used to identify one similar and several dissimilar documents from the corpus data for each query in the query data.

The cross encoder is then used to calculate a similarity between each query and its corresponding (dis)similar documents.

Currently, we support synthesis jobs for data in English or Multilingual, so when choosing a model you can just provide the synthesis_model_en or synthesis_model_multi object which contains the appropriate models for each of these tasks.

Start Synthesis Run#

Now that we have the query and corpus datasets loaded as DocumentArrays, we can begin our synthesis run. We only need to provide the query and corpus data and the models that we are using.

The num_relations parameter is set to 10. This parameter determines how many documents are retrieved for each query. There will always be one similar document and (num_relations - 1) dissimilar documents retrieved. These dissimilar documents are what make up the generated documents, so the size of the generated DocumentArray is always equal to len(query_data) * (num_relations - 1). By default this parameter is set to 10, meaning that the size of the generated dataset would be twice as large as the size of the query dataset.

from finetuner.model import synthesis_model_en

synthesis_run = finetuner.synthesize(


Now that we’ve created a run, we can check its status. You can monitor the run’s progress with the function synthesis_run.status(), and the logs with synthesis_run.logs() or synthesis_run.stream_logs().

Note: The job will take around 15 minutes to finish.

for entry in synthesis_run.stream_logs():

Dependending on the size of the training data, some runs might take up to several hours. You can easily reconnect to your run later to monitor its status.

import finetuner

synthesis_run = finetuner.get_run('my-synthesis-run')
print(f'Run status: {run.status()}')

Retrieving the data#

Once the synthesis run has finished, the synthesised data will be pushed to the Jina AI Cloud under your account. The name of the pushed DocumentArray will be stored in synthesis_run.train_data.

train_data_name = synthesis_run.train_data
train_data = DocumentArray.pull(train_data_name)

Start Training with Synthesised Data#

Using your synthesised data, you can now train a model using the MarginMSELoss function.

We have prepared the index and query datasets xmarket-gpl-eval-queries and xmarket-gpl-eval-queries so that we can evaluate the improvement provided by training on this data:

Note: if you use synthesis_model_multi for training data synthesis on languages other than English, please choose distiluse-base-multi or bert-base-multi as the backbone embedding model.

from finetuner.callback import EvaluationCallback

training_run =

Just as before, you can monitor the progress of your run using training_run.stream_logs():

for entry in training_run.stream_logs():


Our EvaluationCallback during fine-tuning ensures that after each epoch, an evaluation of our model is run. We can access the evaluation results in the logs using print(training_run.logs()):

Training [3/3] ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 470/470 0:00:00 0:02:34  loss: 5.191
INFO     Done                                                                     
DEBUG    Finetuning took 0 days, 0 hours 11 minutes and 55 seconds                  
INFO     Metric: 'precision_at_k' before fine-tuning:  0.16063 after fine-tuning: 0.20824
INFO     Metric: 'recall_at_k' before fine-tuning:  0.29884 after fine-tuning: 0.39044
INFO     Metric: 'f1_score_at_k' before fine-tuning:  0.13671 after fine-tuning: 0.18335
INFO     Metric: 'hit_at_k' before fine-tuning:  0.64277 after fine-tuning: 0.70012 
INFO     Metric: 'average_precision' before fine-tuning:  0.34342 after fine-tuning: 0.41825
INFO     Metric: 'reciprocal_rank' before fine-tuning:  0.40001 after fine-tuning: 0.47258
INFO     Metric: 'dcg_at_k' before fine-tuning:  1.49599 after fine-tuning: 1.89955 
fine-tuning:  1.49618 after fine-tuning: 1.77899
INFO     Building the artifact ...                                                  
INFO     Pushing artifact to Jina AI Cloud ...                                      

The amount of improvement is highly dependent on the amount of data generated during synthesis, as the amount of training data increases, so will the performance of the finetuned model. To increase the number of documents generated, we can either increase the size of the query dataset provided to the finetuner.synthesize function, or increase value of the num_relations parameter, which will result in more documents being generated per query. Conversely, choosing a smaller value for num_relations would result in shorter generation and training times, but less improvement after training.

To better understand the relationship between the amount of training data and the increase in performance, have a look at the how much data? section of our documentation.