English

The WMT Encoder tool converts input text into encoder representations suitable for machine translation research and evaluation. It supports multiple transformer-based variants and subword tokenization to produce stable, fixed-length embeddings that can be used in scoring, similarity measurements, and feature extraction for MT pipelines. Conceptually, text is normalized and tokenized, passed through the selected encoder, and pooled to form a sentence embedding. Users can request per-token embeddings for debugging translation decisions or attention studies, or obtain a single vector for downstream tasks. The tool is language-agnostic in its outputs, but results depend on the chosen variant and tokenization scheme. Unique value lies in configurable pooling methods, batch processing, and optional normalization, enabling researchers to compare encoder configurations with minimal setup. Core features include support for mean, max, and CLS pooling, optional per-token vectors, and JSON-friendly outputs that integrate with evaluation scripts. Use cases include MT model comparison, embedding-based similarity tests, quality estimation, and classroom demonstrations of encoder behavior. The tool abstracts away model hosting details while offering reproducible representations that align with common WMT benchmarks. Practical workflows involve selecting an encoder_variant, choosing a pooling method, and exporting embeddings for downstream scoring or visualization.