Latxa-Qwen3-8B-Literary-v2-ca-eu is a LoRA adapter for Catalan-Basque literary machine translation. It continues fine-tuning from the project general-domain checkpoint and targets bidirectional literary translation for the ca->eu and eu->ca directions.

Model details

Developed by: Paula Guerrero and Iker Gutierrez
Affiliation: University of the Basque Country (EHU)
Model type: LoRA adapter for HiTZ/Latxa-Qwen3-VL-8B-Instruct
Languages: Catalan (ca), Basque (eu)
Domain: Literary translation
Base model: HiTZ/Latxa-Qwen3-VL-8B-Instruct
Continued from: pguerrero-igutierrez/Latxa-Qwen3-8B-General-eu-ca
Repository: pguerrero-igutierrez/Latxa-Qwen3-8B-Literary-v2-ca-eu
Collection: pguerrero-igutierrez/mt-domain-adaptation-ca-eu

Sources

Hugging Face repository: https://huggingface.co/pguerrero-igutierrez/Latxa-Qwen3-8B-Literary-v2-ca-eu
Hugging Face collection: https://huggingface.co/collections/pguerrero-igutierrez/mt-domain-adaptation-ca-eu
Project repository: https://github.com/pguerrero-igutierrez/MT-domain-adaptation
Paper source: https://github.com/pguerrero-igutierrez/MT-domain-adaptation/tree/main/paper

Intended use

This model is intended for research on low-resource Catalan-Basque literary translation, especially in settings where in-domain parallel data is scarce and synthetic back-translation data is used for adaptation.

Supported prompting directions:

eu->ca: Itzuli testu hau euskaratik katalanera:\n\n{source}
ca->eu: Tradueix aquest text del català al basc:\n\n{source}

Out-of-scope use

High-stakes use without human review
Professional literary publishing without post-editing
Medical, legal, or safety-critical translation workflows
General multilingual tasks outside Catalan-Basque translation

Training data

This adapter was trained on the same literary corpora as literaryv1, built through Spanish-pivot synthetic data generation and back-translation:

backtranslated-corpus/ca-literary_trilingual.json
backtranslated-corpus/eu-literary-EhuHac.jsonl

The model was then continued from the project general checkpoint rather than trained directly from the base model.

Training procedure

LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization: 4-bit NF4
Max sequence length: 768
Epochs: 3
Batch size: 4
Gradient accumulation: 8
Learning rate: 5e-5
Scheduler: cosine
Warmup ratio: 0.05
Seed: 42
Checkpoint selection: best validation BLEU

Evaluation

Results on the literary held-out test set:

Direction	chrF++	BLEU	TER	COMET
`eu->ca`	34.51	7.44	87.66	68.72
`ca->eu`	25.87	2.31	100.74	64.34
Overall	30.02	5.17	93.81	65.44

In the project experiments, this continued-adaptation literary model performed slightly below the direct literary SFT model (literaryv1) across the reported literary metrics.

Limitations

Trained on synthetic literary supervision rather than human-translated in-domain CA-EU parallel data
Literary quality aspects such as style, voice, and fluency are only partially captured by automatic metrics
CA->EU literary performance remains challenging, especially under word-level metrics such as BLEU

Usage

This repository contains adapter weights, so it must be loaded on top of the base model.

python

import torch
from peft import PeftModel
from transformers import AutoTokenizer, Qwen3VLForConditionalGeneration

base_id = "HiTZ/Latxa-Qwen3-VL-8B-Instruct"
adapter_id = "pguerrero-igutierrez/Latxa-Qwen3-8B-Literary-v2-ca-eu"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base_model = Qwen3VLForConditionalGeneration.from_pretrained(
    base_id,
    device_map="auto",
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = "Tradueix aquest text del català al basc:\n\nBon vespre."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

If you use this model, please cite the project repository:

bibtex

@misc{guerrero-gutierrez-2026-caeu-mt,
  title        = {Domain Adaptation for Catalan-Basque Machine Translation via Synthetic Data and Continued Fine-Tuning},
  author       = {Guerrero, Paula and Gutierrez, Iker},
  year         = {2026},
  note         = {Unpublished manuscript}
}

Contact

Paula Guerrero: pguerrero005@ikasle.ehu.eus
Iker Gutierrez: igutierrez134@ikasle.ehu.eus