tatooeba-en-pivot-marian

Tatoeba English-Pivot Marian

This repository contains a Marian NMT checkpoint packaged in OPUS-MT style.

Model Details

• Source language: multilingual (English-pivot multilingual)
• Target language: multilingual (English-pivot multilingual)
• Task: translation
• Framework: Marian NMT
• Training data: TartarusXXX/tatooeba-synthetic and TartarusXXX/agentlans-high-quality-english-sentences-synthetic
• Validation set: BOUQuET + held-out Tatoeba (27,000 lines, 54 named subsets)

Evaluation target languages: en, tr, kmr, de, es, fr, el, bg, ru, ka, hy, fa, ar, ckb, ur

Validation

• Best translation: 43.7884
• Epoch: 2
• Update: 192254
• Timestamp: 2026-07-03 07:20:37
• Validation set: BOUQuET + held-out Tatoeba (27,000 lines, 54 named subsets)

Raw best line:

[2026-07-03 07:20:37] [valid] Ep. 2 : Up. 192254 : translation : 43.7884 : new best

Update 192,254 summary (standard corpus BLEU, detokenized SentencePiece output):

Recent validation results:

Files

• tatooeba-en-pivot.spm32k-spm32k.transformer-big.model1.npz.best-translation.npz
• tatooeba-hq-en-pivot.spm32k-spm32k.vocab.yml
• tatooeba-hq-en-pivot.spm32k-spm32k.vocab.yml
• opus.src.spm32k-model
• opus.trg.spm32k-model
• tatooeba-hq.en-pivot.continue.bouquet.transformer.mk
• _tatooeba-hq.common.mk
• tatooeba-en-pivot.spm32k-spm32k.transformer-big.model1.npz.yml
• tatooeba-en-pivot.spm32k-spm32k.transformer-big.train1.log
• tatooeba-en-pivot.spm32k-spm32k.transformer-big.valid1.log
• tatooeba-en-pivot.spm32k-spm32k.transformer-big.valid1.sets.tsv
• tatooeba-en-pivot.spm32k-spm32k.transformer-big.model1.npz.progress.yml
• bouquet+tatooeba-hq-dev.src
• bouquet+tatooeba-hq-dev.trg
• bouquet+tatooeba-hq-dev.src.sets.tsv
• tatooeba-hq-en-pivot.spm32k-spm32k.transformer-big.valid1.192254.out
• infer_tatooeba_en_pivot.py
• best-bleu.decoder.yml
• translate_with_marian.py

Usage

This is a native Marian checkpoint, not a Transformers/PyTorch checkpoint. Install CPU builds of marian-decoder and spm_encode, clone this repository, then translate raw text with the included wrapper:

python infer_tatooeba_en_pivot.py \
  --source-lang en --target-lang tr \
  --text "How are you today?"

For mixed directions, pass TSV rows in the form source_language<TAB>target_language<TAB>text:

python infer_tatooeba_en_pivot.py --tsv requests.tsv

Only English↔N directions are supported. The non-English languages are Turkish (tr), Northern Kurdish (kmr), German (de), Spanish (es), French (fr), Greek (el), Bulgarian (bg), Russian (ru), Georgian (ka), Armenian (hy), Persian (fa), Arabic (ar), Central Kurdish (ckb), and Urdu (ur).

Reproducibility assets

• bouquet+tatooeba-hq-dev.src and .trg: the exact 27,000-line validation pair.
• bouquet+tatooeba-hq-dev.src.sets.tsv: offsets for all 54 named subsets.
• tatooeba-hq-en-pivot...valid1.192254.out: checkpoint hypotheses.
• tatooeba-en-pivot...valid1.sets.tsv: complete named-set score history.
• tatooeba-en-pivot...progress.yml: Marian checkpoint progress metadata.

The decoder YAML uses repository-relative paths. The checkpoint SHA-256 is 4b7c33764d4bb977a81d282e81f4f05c5ca37ab511c27e662f7b3a7e175f8fa9.