Russian Qwen adaptation
According to the model card, this adapts T-lite-it-1.0 to Russian, with hub tags pointing to Qwen/Qwen2.5-7B as base and fine-tune.
Open Source Model Profile · RefalMachine
RuadaptQwen2.5-7B-Lite-Beta
RuadaptQwen2.5-7B-Lite-Beta is a 7.57B-parameter Qwen2 Russian adaptation from RefalMachine. According to the model card, tokenizer replacement plus LEP targets faster Russian generation.
- Publisher
- RefalMachine
- Task
- text-generation
- Model type
- qwen2
- License
- apache-2.0
- Library
- Unknown
- Publication status
- Accepted · not indexed
Model overview
RuadaptQwen2.5-7B-Lite-Beta is published by RefalMachine as a Qwen2-based text-generation model. The captured configuration identifies Qwen2ForCausalLM and Safetensors metadata reports 7566071296 parameters. According to the model card, it is a work-in-progress Russian adaptation with tokenizer replacement, continued pretraining, and LEP, with hub tags pointing to Qwen/Qwen2.5-7B.
Recorded capabilities
Tokenizer replacement and LEP
According to the model card, the tokenizer was replaced, then continued pretraining was followed by LEP Learned Embedding Propagation.
Reported Russian speed gain
According to the model card, an extended tiktoken cl100k plus 48k-token unigram tokenizer raises Russian generation speed by up to 60%.
Qwen2 architecture and size
Captured config reports Qwen2ForCausalLM and qwen2, with Safetensors metadata reporting about 7.57B parameters.
Use cases in the source record
- Russian-language conversational generation where the publisher describes faster output from the extended tokenizer.
- Chat-template inference tests on the recorded 8B BF16 conversational setup described in the captured card text.
Limitations and unknowns
- According to the model card, the record is work in progress at v1, so details may change.
- No evaluation results were extracted from this record.
- No context-window value was extracted from this record.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: RefalMachine/RuadaptQwen2.5-7B-Lite-Beta
Captured: Unknown. Processed: 2026-09-07T19:35:15.207498+00:00.
Описание модели WORK IN PROGRESS!!! Текущая версия v1. Адаптация модели T-lite-it-1.0 на русский язык. В модели был заменен токенизатор, затем произведено дообучение (Continued pretraining) на русскоязычном корпусе, после чего была применена техника LEP (Learned Embedding Propagation). Благодаря новому токенизатору (расширенный tiktoken cl100k с помощью униграм токенизатора на 48 т. токенов) скорость генерации* русскоязычных текстов возрасла до 60% по сравнению с исходной моделью T-lite-it-1.0. *Под скоростью генерации подразумевается количество русскоязычных символов/слов в секунду на одинаковых текстовых последовательностях. Попро…
F001F002F003F004F005F006F007F008F009F010F011F012F013F014