Hugging Face | GitHub | LinkedIn | WhatsApp | Portfolio

omt-en-es-books

A LoRA adapter for Helsinki-NLP/opus-mt-en-es, trained on opus_books (en-es) to adapt an already-strong translator toward literary prose.

This repository holds adapter weights only (2.4 MB). It does nothing without the base model.

On 1,000 held-out sentence pairs it scores 19.99 sacreBLEU, against 17.09 for the untouched base model — a change of +2.90.

The adapter improves on the base model on this domain.

Results

	sacreBLEU	vs base
`Helsinki-NLP/opus-mt-en-es` (untouched)	17.09	—
+ this adapter	19.99	+2.90

Measured on 1,000 pairs from the held-out 20% split, beam size 4, sacreBLEU with default tokenisation.

Epoch	sacreBLEU	vs base
1	19.23	+2.14
2	19.89	+2.80
3	19.99	+2.90

The checkpoint published here is the best epoch by BLEU, not the last one.

Translations side by side

English	Reference	Base	With adapter
Kitty, on the contrary, was more active than usual and even more anima	Kitty, al contrario, estaba más activa a incluso más animada que nunca	Kitty, por el contrario, estaba más activa que de costumbre y aún más	Kitty, por el contrario, estaba más activa que de costumbre y aún más
All necessary preparations shall be made for your return.	Se darán las órdenes necesarias para su regreso.	Se harán todos los preparativos necesarios para su regreso.	Se harán todos los preparativos necesarios para su regreso.
Maintenant partons, allons, allons vers Sion.	¡Demasiado larga ha sido ya la pausa! ¡Adelante!	Los partens, los allons, los allons de Sion.	Los inquilinos son partens, allons, allons vers Sion.
Thursday, Aug.	Jueves 20 de agosto.	Jueves, agosto.	Jueves, agosto.

Usage

The adapter needs its base model:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import PeftModel

tok  = AutoTokenizer.from_pretrained("emonHasan01/omt-en-es-books")
base = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-es")
model = PeftModel.from_pretrained(base, "emonHasan01/omt-en-es-books")

inputs = tok("Hello, how are you today?", return_tensors="pt")
out = model.generate(**inputs, max_length=128, num_beams=4)
print(tok.decode(out[0], skip_special_tokens=True))

To compare against the base model in the same process:

with model.disable_adapter():
    base_output = model.generate(**inputs, max_length=128, num_beams=4)

Merging the adapter in for faster inference:

merged = model.merge_and_unload()
merged.save_pretrained("merged-en-es")

Training


Base model	`Helsinki-NLP/opus-mt-en-es` (79M parameters)
Trainable	589,824 parameters (0.75% of the model)
LoRA rank / alpha / dropout	16 / 32 / 0.05
Target modules	`q_proj`, `v_proj`
Learning rate	0.0003, AdamW, 5% warmup
Batch size	32
Epochs	3, best checkpoint by BLEU
Max length	128 tokens
Training pairs	74,776
Hardware	Tesla T4
Training time	~26 minutes

Two details that matter for Marian

MarianTokenizer carries two SentencePiece models, source.spm and target.spm, and current_spm defaults to the source one. Tokenizing Spanish targets with a plain tokenizer(targets) call segments them with the English model. Targets here were tokenized with text_target=, which switches to the target model.

Labels were left unpadded in preprocessing so DataCollatorForSeq2Seq could pad them with -100. Padding them to max_length with pad_token_id instead would put padding positions into the loss, and at 128 tokens most positions are padding.

Both mistakes train without any error and simply produce a worse model.

Limitations

The base model is already strong. Any adapter has to beat a translator trained on far more parallel text than opus_books contains. Check the comparison above before assuming this one helps your use case.
Literary domain. opus_books is translated literature. Technical, conversational, and contemporary text are out of distribution.
BLEU is a blunt measure. It rewards n-gram overlap with one reference. A fluent, correct translation that words things differently is penalised. Read the side-by-side table, and judge your own domain yourself.
Evaluated on 1,000 pairs, not the full 18,694-pair test split, to keep beam search affordable. Expect noise of a few tenths of a BLEU point.
Adapter only. It will not load on its own, and it will not work on a different base model.

Reproducing

Training, the baseline measurement, both charts and this card come from one Colab notebook on a Tesla T4. The raw numbers are in results.json.

Citation

@inproceedings{tiedemann2020opusmt,
  title     = {OPUS-MT — Building open translation services for the World},
  author    = {Tiedemann, J{\"o}rg and Thottingal, Santhosh},
  booktitle = {Proceedings of EAMT},
  year      = {2020}
}