Yehor/kulyk-en-uk
Yehor/kulyk-en-uk is a lightweight 354 million parameter machine translation model specifically designed for translating English to Ukrainian. Fine-tuned on 40 million high-quality samples, it achieves a BLEU score of 27.24 on the FLORES-200 benchmark. This model is optimized for efficient deployment, requiring only 1 GB of RAM, and offers high token throughput for practical English-to-Ukrainian translation tasks.
Loading preview...
Model Overview
Yehor/kulyk-en-uk is a compact 354 million parameter model developed by Yehor for efficient machine translation from English to Ukrainian. It is based on the LFM2 model architecture and was fine-tuned using 40 million quality-filtered samples over 1.4 epochs.
Key Capabilities & Performance
- English to Ukrainian Translation: Specialized for this specific language pair.
- High Efficiency: Requires only 1 GB of RAM for operation (with bf16 precision).
- Strong Throughput: Achieves 229.93 tokens per second (batch size 1) and up to 8392.48 tokens per second (batch size 64).
- Benchmark Performance: Attains a BLEU score of 27.24 on the FLORES-200 dataset, indicating solid translation quality for its size.
- Context Length: Supports a maximum sequence length of 2048 tokens during training.
Training Details
The model was trained with a learning rate of 3e-5, using a cosine learning rate scheduler and a warmup ratio of 0.05. Training utilized Flash Attention 2 and DeepSpeed for optimization, running on two NVIDIA RTX 3090 Ti GPUs. The model's name honors Sergiy Kulyk, a Ukrainian diplomat.
Use Cases
This model is ideal for applications requiring fast and resource-efficient English-to-Ukrainian translation, particularly in environments with limited computational resources. A demo is available for testing.