opus-mt-en-ru

Performs bidirectional sequence-to-sequence translation from English to Russian using the Marian NMT framework, a PyTorch-based encoder-decoder architecture with multi-head attention and learned positional embeddings. The model was trained on parallel corpora from the OPUS project and supports both PyTorch and TensorFlow inference backends, enabling deployment across heterogeneous environments (CPU, GPU, TPU). Tokenization uses SentencePiece subword segmentation for handling morphologically rich Russian and productive English compounds.

Supports multi-sentence and document-level translation via batched inference with configurable beam search (width 1-5), length penalties, and sampling-based decoding. The model's generate() method accepts batch inputs of variable length, automatically pads sequences to the longest in the batch, and applies length normalization to prevent bias toward shorter translations. Beam search explores multiple hypotheses in parallel, enabling trade-offs between translation quality and latency.

Model weights are serialized in HuggingFace safetensors format and compatible with PyTorch (.pt), TensorFlow (.pb), and ONNX Runtime backends, enabling deployment across diverse inference stacks without retraining. The transformers library automatically handles format conversion and backend selection at load time. Supports deployment on Azure ML, AWS SageMaker, and self-hosted Kubernetes clusters via standard container images.

Uses SentencePiece BPE (Byte-Pair Encoding) tokenization trained on parallel English-Russian corpora, enabling efficient handling of morphologically rich Russian (case, gender, aspect inflections) and productive English compounds. The tokenizer learns ~32K subword units that balance vocabulary coverage with sequence length, reducing OOV (out-of-vocabulary) rates compared to word-level tokenization. Supports reversible detokenization for reconstructing original text from token sequences.

The pre-trained Marian encoder-decoder can be fine-tuned on domain-specific parallel corpora using standard PyTorch training loops or HuggingFace Trainer API, enabling rapid adaptation to specialized vocabularies and translation patterns. Fine-tuning leverages the model's learned representations from OPUS pre-training, requiring only 10K-100K parallel sentences to achieve significant quality improvements on target domains. Supports parameter-efficient fine-tuning via LoRA (Low-Rank Adaptation) to reduce memory overhead and training time.