Building large-scale English-Romanian literary translation resources with open models

Area of Science:

• Natural Language Processing
• Machine Translation
• Computational Linguistics

Background:

• Literary translation presents unique challenges for machine translation, especially for low-resource languages like Romanian.
• Small open-weight models have not yet been effectively applied to literary translation tasks.

Purpose of the Study:

• To introduce the TinyFabulist Translation Framework (TF2) for English to Romanian literary translation.
• To develop and evaluate an open-weight model for high-quality literary translation in low-resource settings.

Main Methods:

• Utilized a large dataset of synthetic English fables (DS-TF1-EN-3M) to generate 15,000 Romanian references using a high-performing large language model (LLM).
• Applied a two-stage fine-tuning process to a 12-billion parameter open-weight model: instruction tuning for narrative style and adapter compression for efficiency.
• Evaluated translation quality using a five-dimension LLM-based rubric and Bilingual Evaluation Understudy (BLEU) scores.

Main Results:

• The fine-tuned model (TF2-12B) demonstrates strong fluency and adequacy, significantly reducing the performance gap with proprietary models.
• The open-access model and datasets (DS-TF2-EN-RO-3M, DS-TF2-EN-RO-15K) are publicly released.
• Achieved a cost-effective solution for literary translation in low-resource scenarios.

Conclusions:

• The TinyFabulist Translation Framework (TF2) offers a reproducible pipeline for cost-efficient literary translation and cross-lingual narrative generation.
• TF2 facilitates the adoption of open models for culturally significant content in low-resource languages.
• This work advances research in machine translation for literary domains and low-resource languages.