8 billion parameter data update

One of the famous Persian language model groups developed by Paret Artificial Intelligence Research Center is called Derna. One of the most widely used LLMs of this group is their large 8 billion parameter language model, which has been released as open source; This model has been welcomed by developers due to its accurate performance in Persian language.

The second edition is worth 8 billion

Now this language model has been upgraded to the second version and several new features have been added to it and its overall performance has been significantly improved.

It is interesting to know that in recent months, large Persian language models have been able to establish their place in the artificial intelligence ecosystem of the country and be used in many smart tools. Farsi language users have also established an effective relationship with these tools, and most of the high-quality artificial intelligence services have appeared successful in attracting the attention of the audience.

This success is due to infrastructures such as LLMs, which have been made available for free and open source to domestic developers and provided them with the opportunity to compete with prominent foreign services.

What updates has the big language model of Derna received?

As it was said, the big language model of 8 billion parameters from the Derna group, in its big update to version 2, has experienced significant improvements. Among the most important upgrades of the new version, we can mention the increase in the length of the input data (Context length), which with a leap It has increased 16 times, from 8 thousand to 128 thousand tokens. This dramatic growth in the number of tokens helps the model to accept longer inputs and provide more accurate answers to user queries.

In addition to what has been said, developers can connect third-party services and various APIs to their product using this enhanced model and make a wide range of services available to their users as a unit, for example, a The intelligent chatbot can suggest the most appropriate option based on the price and travel date of the user using the API of the airline ticket sales centers and, if approved, book the desired ticket instantly.

The large language model with 8 billion parameters has also been evaluated in the Persian LLM Leaderboard Part benchmark, and its performance results can now be viewed and reviewed on the HuggingFace page.

Part’s special attention to personal developers and start-up businesses

Besides offering large-scale models for professional and organizational uses, Part also pays special attention to personal developers and start-up businesses. These developers do not have access to advanced and expensive hardware and need a tool that can be used with personal systems.

In this regard, Part has released a new member of the Tuka model language group under the name of Model (SBert) as open source to enable the development of intelligent chatbots for a wide range of developers.

Toka can be implemented in different systems and it has gained this capability due to its low volume. All chatbot services, regardless of the form of their activity or the use of internal and external models, need to use this type of language models; Now, thanks to Tuka, this model can be used locally and implemented with limited resources and hardware.

That being said, Toka is a good choice for RAGs and FAQ chatbots.