Ksenia-Dmitrieva/Corpus-D-opus-mt-en-fr-finetuned-en-to-fr

This model was fine-tuned on a modified corpus of English-French chat translation texts from WMT24 Chat Shared Task.

Original model: Helsinki-NLP/opus-mt-en-fr

Training corpus: parallel training corpus of WMT texts with artificially created errors: 45% of the segments in the source text column in the corpus were replaced with translations into a language (Russian) other than source language.

Training parameters:

training_args = Seq2SeqTrainingArguments(

f"TEST-{modelname}-finetuned-{sourcelang}-to-{target_lang}",

report_to="none",

eval_strategy = "epoch",

save_strategy = "epoch"

learning_rate=1.278740919668083e-06,

perdevicetrainbatchsize=2,

perdeviceevalbatchsize=2,

weight_decay=0.07280588826629236,

numtrainepochs=3,

warmup_steps=499

)

Model quality

Estimated average quality from model's translations of test set

BLEU: 47.9

ChrF++: 72.8

COMET: 90.5

COMET Kiwi: 84.3