Enhancing LLM Translation Performance for Spanish–Valencian through Supervised Fine-Tuning and Reinforcement Learning

Paula Guerrero Castelló


Abstract
Valencian, the Western Catalan variety used in the Valencian Community of Spain, lacks a dedicated language code in most multilingual machine translation (MT) systems, and is systematically rendered closer to the standard written Eastern Catalan used in Catalonia. We address this gap by adapting TranslateGemma-4B-IT, a 4-billion-parameter instruction-tuned (IT) large language model (LLM) specialized for translation, via three post-training strategies using only public corpora and Quantized Low-Rank Adaptation (QLoRA): (i) supervised fine-tuning (SFT); (ii) Group Relative Policy Optimization (GRPO), a reinforcement learning (RL) technique, with chrF plus a naturalness reward (GRPOV1); and (iii) GRPO with a composite automatic-metric reward (GRPOV2). Our results suggest that reward-function alignment with the target dialect is a key determinant of RL success in low-resource dialectal MT.
Anthology ID:
2026.eamt-2.27
Volume:
Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2)
Month:
June
Year:
2026
Address:
Tilburg, The Netherlands
Editors:
Dimitar Shterionov, Eva Vanmassenhove, Mirella De Sisto, Fred Blain, Javad Pourmostafa Roshan Sharami, Lisa Lepp, Chiara Manna, Argentina Anna Rescigno, Alina Karakanta, Ayla Rigouts Terryn, Manuel Lardelli, Natalia Resende, Elena Murgolo, Janiça Hackenbuchner, Anna Zaretskaya, Miquel Esplà-Gomis, Thierry Etchegoyhen, Dagmar Gromann, Rachel Bawden, Barry Haddow, Sara Szoc, Mikel Forcada, Helena Moniz
Venue:
EAMT
SIG:
Publisher:
European Association for Machine Translation
Note:
Pages:
72–79
Language:
URL:
https://aclanthology.org/2026.eamt-2.27/
DOI:
Bibkey:
Cite (ACL):
Paula Guerrero Castelló. 2026. Enhancing LLM Translation Performance for Spanish–Valencian through Supervised Fine-Tuning and Reinforcement Learning. In Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2), pages 72–79, Tilburg, The Netherlands. European Association for Machine Translation.
Cite (Informal):
Enhancing LLM Translation Performance for Spanish–Valencian through Supervised Fine-Tuning and Reinforcement Learning (Castelló, EAMT 2026)
Copy Citation:
PDF:
https://aclanthology.org/2026.eamt-2.27.pdf