What are Super Models and how do they work?

The Text Titan, the English Elder, the Dutch Dean: Super Models are one of the most important developments in Transkribus. Why? Because Super Models produce more accurate transcriptions of diverse collections of documents, without the need to train your own model.

In this post, we will take a closer look at Super Models, the technology behind them, and give an overview of the Super Models that you can use to get the best possible transcriptions with Transkribus.

Key Takeaways:

- Super Models are capable of producing accurate transcriptions "out of the box", even with diverse collections of documents.

- They achieve this by using advanced transformer technology and much larger, generic datasets.

- There are Super Models available for various different languages, as well as the Text Titan for all Latin Scripts. 

Each Super Model has its own distinct name and avatar. © Transkribus

What is a Super Model?

The majority of models available on Transkribus, including those you can train yourself, are built using the open-source PyLaia text recognition engine. PyLaia's strength lies in creating highly specialised models tailored to specific document types. For instance, if you have a collection containing handwritten Dutch documents from the 18th century, then a model trained on 18th-century Dutch handwriting should be able to produce an accurate transcription of your documents. This specificity allows for impressive accuracy when the training data closely matches your documents.

However, this very specificity becomes a limitation when dealing with diverse collections. If your archive contains documents in multiple languages, in various handwriting styles, or even with a mix of handwritten and printed text, training individual PyLaia models for each segment can be a time-consuming and resource-intensive process.

This is where our Super Models can help. Designed to handle much more heterogeneous collections, these powerful models deliver accurate transcriptions across numerous languages and handwriting styles without requiring any additional training. Let's take a look at how they achieve this.

The German Genius is a Super Model tailored specifically for the German language. © Transkribus

How are Super Models created?

The main difference with Super Models is that they are trained with advanced transformer technology. This kind of technology is much more complex than the technology behind the PyLaia models, and so training Super Models requires specialist knowledge as well as powerful computers. 

Super Models are trained in three steps. Firstly, the model is trained with a very, very large dataset of mainly generic text. This data includes handwritten and printed texts in many different languages, giving the model a comprehensive foundation in analysing images and extracting text from them.

But because this initial training data is generic and not necessarily historical, the next step is to fine-tune the model and make it more appropriate for historical documents. This involves changing some minor details in the model, as well as training it with a large dataset of generic historical data, improving its ability to read these kinds of documents.

This base Super Model can then be tweaked even further by training it with even more specialised datasets. For example, for our Dutch Dean Super Model, we added a very large dataset of Dutch historical documents to the model, increasing its accuracy when transcribing Dutch documents.

The Faucon Français has a CER of 6.6% and is suitable for a wide range of French-language documents. © Transkribus

How is this different to PyLaia models?

In many ways, training a Super Model isn't that different to training a PyLaia model. Both involve feeding the model a large amount of training data, so that it can learn how to transcribe the various handwriting styles and typefaces contained in that data.

With a Super Model, though, most of the training data is very generic. Only a small proportion of the entire training dataset is dedicated to fine-tuning the model to read historical documents. In contrast, a PyLaia model is trained with very specific data from the start. This means that they will always be biased towards the specialist end of the spectrum, whereas Super Models have a much stronger "basic reading" competence, making them suitable for a much wider range of documents.

The English Elder was trained on a very diverse collection of English-language documents from around the world. © Transkribus

What kind of documents are Super Models good for?

Super Models are ideal for very generic collections containing many different languages and hands. This is also because they are based on more powerful technology and trained with much more data, and so are better at multitasking. A PyLaia model can only transcribe French handwriting from the 17th century, whereas a Super Model can, for example, transcribe French handwriting from the 17th century, German Fraktur typeface, and Italian medieval manuscripts all at the same time.

Super Models are also intended to be used "out of the box" without the need for further training or tweaking. This makes them a good choice if you don't have the time or resources to train your own model.

What kind of documents aren't Super Models good for?

As you can probably imagine, if you have a very specific collection, then it may still be worth using or training a PyLaia model for it.

To give an example, we recently had a Transkribus project with a training set of 15,000 pages of very homogeneous index cards. We created both a specially trained Super Model and a PyLaia model based on this data. But despite the Super Model's superior technology and knowledge, the PyLaia model still outperformed it, as the PyLaia model's focused training made it the "expert" for that particular material. In other words, for highly specialised collections with lots of representative data, a custom PyLaia model remains a strong option.

The Dansk Dokumentalist was created with the help of Johan Heinsen from Aarlborg University. © Transkribus

Which Super Models are available?

Our most comprehensive Super Model is the Text Titan. This powerful model has been trained on all the major languages in the Latin script, including German, French, English, Dutch, and many more. It is, therefore, the go-to model for multilingual collections, and usually returns very good results on monolingual collections too. We are constantly updating the Text Titan with more training data. The latest version is the Text Titan I bis, which can be tried here.

• Dutch Dean
• German Genius
• Dansk Dokumentalist
• Polski Bizon
• Faucon Français
  
• English Elder
• Spanish Sage

We also have several new Super Models in the pipeline, keep checking our BlueSky or LinkedIn pages for updates.

The Text Titan is our most comprehensive Super Model and can be used for historical documents written in the Latin script. © Transkribus

Can I train my own Super Models?

Unfortunately not. Super Models require specialist knowledge and powerful computers to train, and so this is not possible for users to do individually within Transkribus.

Can I use Super Models as a base model?

Also not. Because Super Models are based on different technology, they cannot be used as base models when training a PyLaia model.

However, one of the big advantages of Super Models is that their accuracy with generic collections makes them excellent for creating Ground Truth. Instead of transcribing your training data manually from scratch, you can use a Super Model to automatically transcribe them and then edit those transcriptions, speeding up the process greatly.

Who can use Super Models?

Operating Super Models requires a lot more computational power, therefore Super Models are only available to subscribers on our paid Scholar, Team, or Organisation plans. 

However, if you want to test our Super Models and see if they are right for your documents, then we offer a free 7-day trial of the Scholar plan, giving you temporary access to Super Models, as well as a host of other advanced tools, such as Table Models and Transkribus Sites. You can activate the free trial by registering for Transkribus and clicking on the "Free Trial" banner in the app.