OPUS-MT-train/tatoeba/TODO.md

Pivot-based data augmentation

• pbt: pivot-back-translated (e.g. translate eng to swe to change fin-eng into fin-swe training data)
• pft: pivot-forward-translated (e.g. translate eng to fin to change eng-swe into fin-swe training data)
• pivotalign: merge fin-eng and eng-swe into fin-swe training data by simply matching eng sentences (see OPUS)
• pftonly pbtonly paonly: targets to use only those pivot-based data sets (no standard train data) (--> distillation)

Cleanup and documentation

• cleanup / remove complicated testset evaluation for multilingual models
• cleanup recipes for tatoeba model training
• improve documentation
• tutorials

Evaluation

• better link information about evaluation of released models (see eval subdir)
• integrate NMT map (move to OPUS-MT-map)
• better score tables / leaderboards for released NMT models

Knowledge distillation and compact models

• systematically test model architectures
• multilingual distillation
• better data selection for student models
• quantization and tuned int8 models (train alphas) - see browsermt/students
• lexical shortlists
• better integration into translateLocally / OPUS-MT-app