Sber has released the source weights of FRIDA-Decisions — a model for making structured decisions based on text. It receives text and predefined answer options, after which it can select the desired category, evaluate the text on a scale, determine a “yes” or “no” answer, or rank several options. The project was developed by the UESMO and SberAI teams and published under the MIT license.
FRIDA-Decisions is based on the FRIDA model, which uses a T5 encoder. The new version contains 823 million parameters and supports four types of tasks: selecting one option, ordinal evaluation, “yes/no” assertion checking, and ranking multiple candidates. The result is a selected option, an evaluation, or an order of given options, not generated text.
The mechanism of FRIDA-Decisions differs from generative language models. In a typical LLM, the answer is formed sequentially, token by token. FRIDA-Decisions encodes the source text and matches it with the given options. This allows the same text to be used for multiple tasks, and when re-processed, to save the intermediate state using a cache.
The developers separately provide a test with a catalog of 243 categories. According to their data, on an RTX 5060 Ti graphics card, processing such a query took about 0.44 seconds when using option packing and state caching. With sequential processing of each option as a separate query, the result was about 4.65 seconds.
The model is also adapted for operation on central processing units. For this, the developers published a version in ONNX format with 8-bit quantization. In one test, a query with text of approximately 384 tokens and three questions was processed in 0.88 seconds on six CPU threads. The PyTorch version in FP32 format under the same conditions showed a result of about 2.28 seconds.
FRIDA-Decisions was also tested on the razvilka dataset, which includes 735 Russian-language examples and 15 types of tasks. These included classification, routing, moderation, sentiment analysis, and ranking. According to the developers' test results, FRIDA-Decisions achieved a metric value of 0.893. For the commercial system TypeSafe Jev, the same comparison showed 0.897, and the FRIDA-Decisions version in ONNX format on CPU showed 0.891.
1.42 million examples and about 1.5 million questions covering 151 types of tasks were used to train the model. According to the developers, 71% of the training data was in Russian, and another 29% was in English. About a quarter of the dataset, according to the team's assessment, contained texts or labels created using language models.