ApsuaTranslator-Lite-v1

Русская версия

ApsuaTranslator-Lite-v1 is a bidirectional Russian–Abkhaz model distilled from ApsuaTranslator-v1. Compared with the full model, its weight file is 6× smaller on disk, loaded weights use 3× less memory, and translation is 2.5× faster. It was developed as part of the Abkhaz Language Corpus project.

The model was trained using Apsua Corpus, a Russian–Abkhaz parallel corpus built through human-in-the-loop translation and editorial review.

Model

Property	Value
Translation directions	Russian ↔ Abkhaz
Parameters	82,152,000
Weights	INT8 Safetensors
Inference precision	FP32 on CPU; FP16 on CUDA/MPS
Maximum input length	128 tokens
Maximum generated length	128 tokens

Evaluation

Case-insensitive BLEU on an independent test set spanning conversation, reviews, informational texts, and literature.

Model	Russian → Abkhaz	Abkhaz → Russian
ApsuaTranslator-v1	25.13	34.60
ApsuaTranslator-Lite-v1	23.57	29.17
ApsuaTranslator-v0	21.94	28.72
Google Translate	20.12	30.74

Usage

pip install "git+https://huggingface.co/apsua/ApsuaTranslator-Lite-v1"

from apsua_translator import Translator

translator = Translator.from_pretrained("apsua/ApsuaTranslator-Lite-v1")

print(translator.translate("Добрый день!", source_lang="ru", target_lang="ab"))
print(translator.translate("Мшыбзиа!", source_lang="ab", target_lang="ru"))

translate accepts one string or a list of strings. Device selection is automatic in the order CUDA, Apple MPS, CPU; it can also be set explicitly.

The release uses the included apsua_translator.Translator runtime and is not a transformers.AutoModel package.

License

Model assets: CC BY 4.0. Runtime code: Apache-2.0.

Project

The project was initiated by Team Abkhazia (Команда Абхазии), with support from the Administration of the President of the Republic of Abkhazia, the Repatriation Fund of the Republic of Abkhazia, and the Presidential Grants Foundation of the Russian Federation.