T-Lite (Large Language Model, LLM)

The main articles are:

• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding

2024

Open access to T-Pro models and updated T-Lite

The T-Technologies Group has opened access to two large language models (LLM) - a T-Pro with 32 billion parameters and an updated T-Lite with 7 billion parameters. Now any Russian company will be able to use their capabilities for free. Representatives of the group told TAdviser on December 11, 2024. According to T-Technologies, as shown by industrial benchmarks (MERA, ruMMLU, Ru Arena Hard, MT Bench and AlpacaEval), T-Pro and T-Lite "surpass all Russian and foreign models in terms of general level of knowledge, ability to conduct dialogue and perform practical tasks."

Open access involves completely free use of models to solve internal problems and create new products based on LLM - regardless of business size. For example, with the help of LLM, you can create clever chat boats in support that are more close to human communication and respond not by scripts, but in live dialogue mode. This will allow you to partially or completely automate the analysis of client cases. Also, with the help of LLM, a business can create assistants for its employees: tools for automatic code writing, reporting, writing research.

The use of models from T-Technologies will allow business to:

• not to create your own multibillion-dollar models from scratch, but to further study the existing base for your needs;
• save on fees to third-party service providers for using their proprietary (closed) models.

Both models (T-Pro and T-Lite) are included in Gen-T - a family of their own specialized language models of the T-Technologies group. Continuous Pretraining technology is used to create models. This is a process in which a model already trained on large amounts of information continues to be trained on materials specific to a specific task or area, and adapts it into Russian. The T-Lite and T-Pro models are based on models of the Qwen-2.5 family, but show higher quality on the tasks of the Russian language than the original models, the company noted. This approach allows T-Technologies to reduce the cost of creating large language models.

"With our business volume (more than 46 million customers and 90 thousand employees), there are tasks of such a level of complexity for which only our own technologies are suitable. When we began to develop products based on large language models - for example, copilots for employees and the AI Assistant Universe - we were once again convinced that the solutions on the market do not meet our requirements, "said Viktor Tarnavsky, Director of Artificial Intelligence at T-Bank. "So we started developing Gen-T, a family of specialized language models."

According to him, the bank's goal was to create solutions of the proper quality and at the same time minimize costs using the developments of the world scientific and engineering community. Having made sure of the effectiveness of the created solution, T-Bank decided to share it with the entire industry and change the approach to using LLM.

Why create your own expensive models if you can take the best from the market and customize it for yourself. Other companies will be able to take over our experience, and the use of LLM will become much wider, "added Viktor Tarnavsky.

Updated T-Lite

The updated version of T-Lite with 7 billion parameters is suitable for further training for specific business tasks. Among the main improvements, according to the developers:

• Accuracy and contextuality. Better captures context, understands complex queries, and produces more accurate answers.
• Generation depth. Creates texts that are closer to the human style of writing, with fewer errors and more logical connectivity.
• Adaptability. The model is easily completed for the needs of specific industries - from finance and medicine to retail and education.

𝓲

"T-Bank"

According to T-Technologies, T-Lite became the best in the category of open models up to 10 billion parameters according to the results of industrial benchmarks, including MERA, ruMMLU, Ru Arena Hard, MT Bench and AlpacaEval.

𝓲

"T-Bank"

T-Pro

Compared to the predecessor T-Lite, the T-Pro model has increased the number of parameters - from 7 to 32 billion. This makes the model more powerful and productive. More parameters allow the model to take into account more context and language features, better memorize information, and draw more accurate and complex conclusions.

𝓲

"T-Bank"

The model works in two modes: it can be further trained for specific business tasks (Fine-tuning), as well as used in prompting mode - to set tasks for the model in dialog mode.

𝓲

"T-Bank"

According to T-Technologies, among the open models of its weight category, T-Pro ranks first in solving problems in Russian according to the industrial benchmarks MERA, ruMMLU, Ru Arena Hard, MT Bench and AlpacaEval. Among proprietary (closed) models, the T-Pro ranks second in a number of benchmarks, behind only GPT-4o.

Availability

The T-Lite and T-Pro models are already available for download on the Hugging Face platform under the Apache 2.0 license.

When used together with the open Turbo Alignment library, companies will be able not to develop AI applications from scratch, but to use ready-made tools.

Product Announcement

In July 2024, T-Bank announced the release of the most powerful Russian language model T-lite. It is designed to create AI solutions in the field of data analysis, search and development of chatbots.

T-lite has 8 billion. Parameters are numerical values that the model adjusts to better understand and generate text. The more parameters, the greater the ability of the model to perform complex tasks, but as the size increases, the cost effectiveness of the model also deteriorates. T-lite, after further training for specific business tasks in the field of natural language processing (NLP), gives a quality comparable to proprietary models with a size of 20 billion parameters, but at the same time several times cheaper to operate, the bank said.

They also noted that on industrial and internal benchmarks the model overtook foreign llama3-8b-instruct and chat-gpt 3.5 in terms of indicators. At the same time, T-lite is created using only 3% of the computing resources that are usually required for this type of model.

T-lite is part of Gen-T - a family of its own specialized language models of T-Bank, which can learn to solve specific highly specialized problems. Unlike universal models such as ChatGPT, Gen-T technology is focused on specific areas and offers solutions with maximum adaptation to the needs of the user, the bank said in a statement.

Our Gen-T models have shown their effectiveness in our services, are optimal in terms of the ratio of used capacities and quality. And we are ready to share this development with other companies, with users, with the professional community. This is our contribution to the development of artificial intelligence in Russia, "said Artem Bondar, head of NLP at the T-Bank Center for Artificial Intelligence.