Dilmaç 28M
A compact model for English-to-Turkish translation
2 min readupdated
The new language model Dilmaç 28M was created specifically for fast text translation from English to Turkish. With a size of just 28 million parameters (weighing about 112 MB in fp32 format or 56 MB in bf16), the model runs entirely locally and can efficiently operate even on standard laptop CPUs. The architecture is an encoder-decoder Transformer with 8 encoder layers and 2 decoder layers, optimized for sentences up to 128 tokens long. The development targets developers and researchers who need a fast and lightweight machine translation tool without relying on cloud APIs.
Key features
- Efficiency: High operating speed on CPUs and integrated graphics thanks to its small footprint and optimized architecture.
- Translation reliability: Shows stable results on standard benchmarks such as FLORES-200.
- Local execution: Runs completely offline without sending data to external servers.
- Flexible decoding configuration: Supports beam search parameters to control quality and prevent looping.
What you can do
- App integration: Add fast text translation to mobile or desktop applications for offline operation.
- Utility creation: Use the model in local bots or scripts to translate messages and documents.
- Draft translation: Generate preliminary translation variants of articles and letters for subsequent human editing.
- Research: Apply as a fast baseline model for distillation and machine translation experiments.
Example
from transformers import AutoModel, AutoTokenizer
repo = "edededdy/dilmac-en-tr-28m"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True)
model.translate("Where is the nearest train station?", tok)
# 'En yakın tren istasyonu nerede?'
Read more
DetailsMetrics: 54.83 chrF++ / 26.32 BLEU on FLORES-200 devtest (with num_beams=4 parameter).
Hardware requirements: Compatible with CPU and GPU, supports optimization down to bf16 format for memory savings and speedup.
Limitations: Works in only one direction (English to Turkish), processes sentences up to 128 tokens long, struggles with headlines and complex contexts.
License: Apache-2.0.










Comments