Researchers Build Framework for Translating Literary Works into Romanian Using Open-Source Models

Researchers have developed the TinyFabulist Translation Framework (TF2), a comprehensive pipeline for creating high-quality literary translations from English into Romanian using openly available machine learning models. The work addresses a significant gap in machine translation research: the challenge of translating narrative fiction into languages with fewer digital resources, particularly using smaller open-weight models rather than expensive proprietary systems.

The Challenge#

Literary translation presents unique demands compared to generic machine translation. It requires capturing not just accurate word-by-word meaning but also narrative voice, stylistic nuance, cultural context, and emotional resonance. Most existing translation research focuses on technical or news-domain content, leaving literary works—fables, novels, and other narrative forms—understudied. Romanian presents an additional hurdle as a lower-resource language with limited parallel translation datasets.

The TF2 Approach#

The framework builds on a foundation of 3 million synthetic English fables (DS-TF1-EN-3M), the largest synthetic fable collection created to date. The researchers generated 15,000 high-quality Romanian translations from this pool using a high-performing large language model, then applied a specialized two-step fine-tuning process to a 12-billion-parameter open-weight model:

• Instruction tuning to teach the model narrative genre conventions and literary style
• Adapter compression to reduce computational overhead while maintaining performance

Evaluation Methodology#

The team assessed results using a multi-faceted rubric covering accuracy, fluency, coherence, stylistic consistency, and cultural adaptation—moving beyond simple automated metrics like BLEU scores toward more nuanced literary evaluation. This combination of detailed linguistic assessment and automated benchmarking provided a more complete picture of translation quality.

Results and Impact#

The fine-tuned TF2-12B model achieved strong performance across fluency and adequacy metrics, approaching the quality of top proprietary translation services while remaining open-source and significantly cheaper to operate. The researchers publicly released the fine-tuned model, two large-scale parallel datasets (3 million and 15,000 sentence pairs), and all supporting code and evaluation scripts, enabling reproducible research and broader adoption of open models for literary translation in language-scarce settings.