synthetic-n2en-marian-transformers

synthetic-n2en-marian-transformers

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

Model Details

• Source language: multi (Multilingual)
• Target language: en (English)
• Task: translation
• Framework: Marian NMT
• Training data: TartarusXXX/synthetic-parallel-16lang-1-4m-gemini-3-1-flash-lite
• Validation set: bouquet-dev-n2en

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

Validation

• Best bleu: 28.8111
• Epoch: 9
• Update: 318000
• Timestamp: 2026-06-17 05:14:00
• Validation set: bouquet-dev-n2en

Raw best line:

[2026-06-17 05:14:00] [valid] Ep. 9 : Up. 318000 : bleu : 28.8111 : new best

Recent validation results:

Files

• synthetic_parallel_16lang_1_4m_gemini_3_1_flash_lite.spm32k-spm32k.transformer.model1.npz.best-bleu.npz
• synthetic_parallel_16lang_1_4m_gemini_3_1_flash_lite.spm32k-spm32k.vocab.yml
• opus.trg.spm32k-model
• opus.src.spm32k-model
• hf-csv-synthetic.n2en.bouquet.transformer.mk
• synthetic_parallel_16lang_1_4m_gemini_3_1_flash_lite.spm32k-spm32k.transformer.model1.npz.yml
• synthetic_parallel_16lang_1_4m_gemini_3_1_flash_lite.spm32k-spm32k.transformer.train1.log
• synthetic_parallel_16lang_1_4m_gemini_3_1_flash_lite.spm32k-spm32k.transformer.valid1.log
• bouquet-dev-n2en.best-bleu.validation-style-spm.tsv
• bouquet-dev-n2en.best-bleu.validation-style-spm.md
• best-bleu.decoder.yml
• translate_with_marian.py

Usage

Install Marian and run:

python translate_with_marian.py --decoder best-bleu.decoder.yml < input.txt > output.txt

The decoder YAML uses paths relative to this repository.

Conversion

Converted from the Marian checkpoint in TartarusXXX/synthetic-n2en-marian using the Transformers Marian converter.

• Source language: multi
• Target languages: en

Transformers Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = "TartarusXXX/synthetic-n2en-marian-transformers"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# This model was trained as `multi` -> `en`.
# If the Marian training data uses target-language tags, include the desired
# target tag in the source sentence, following the training convention.
inputs = tokenizer("Hello world!", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])