proxectonos/Llama-Carvalho-PT-GL_300k

This repository contains proxectonos/Llama-Carvalho-PT-GL_300k, a LoRA adapter for machine translation from Spanish to Galician, built on top of Nos-PT/Llama-Carvalho-PT-GL.

Why this model exists

This checkpoint is the 300k training-size model from our data-scaling experiments on low-resource machine translation with LLM fine-tuning. It is being published because it was the best-performing model for the es-gl direction among the four training sizes we evaluated: 30k, 300k, 600k, and 3.6M parallel sentence pairs.

The motivation for releasing this model is not just to provide another translation adapter, but to document a central experimental result: for Spanish-to-Galician, a closely related language pair, the best translation quality was obtained with a medium-sized corpus rather than with the largest available one. In our experiments, the 300k model achieved the strongest overall results for es-gl, especially in BLEU and human evaluation.

This makes the model an important counterpart to the 3.6M English-Galician checkpoint from the same study. Together, these releases illustrate the main conclusion of the work: the optimal fine-tuning size depends on linguistic distance, and more data is not always better for closely related language pairs.

Model details

• Base model: Nos-PT/Llama-Carvalho-PT-GL
• Model type: LoRA adapter for causal language modeling used as machine translation
• Task: Machine translation
• Direction: Spanish -> Galician
• Languages: Spanish, Galician
• Training method: Supervised Fine-Tuning with LoRA
• Training size: 300,000 parallel sentence pairs

Training data

The adapter was trained on 300k Spanish-Galician parallel sentence pairs.

The training data comes from the CorpusNos MT dataset and includes only human-written text. The corpus covers several text types, including:

• formal institutional language
• scientific text
• spoken-domain material
• multi-word expressions

Training setup

All experiments in the paper used the same base model and the same fine-tuning setup so that differences in quality could be attributed to data scale rather than to architecture changes.

• Base model: Llama-Carvalho-PT-GL
• Fine-tuning strategy: LoRA
• Learning rate: 5e-6
• Scheduler: cosine
• Warmup steps: 150
• Weight decay: 0.01
• Per-device train batch size: 16
• Gradient accumulation steps: 4
• Precision: bf16

Evaluation summary

This model was evaluated in the paper against other checkpoints trained on smaller and larger Spanish-Galician corpora (30k, 600k, 3.6M).

Among the es-gl models, this 300k checkpoint obtained the best overall results:

• Average BLEU across 6 benchmark test sets: 60.63
• Average COMET across 6 benchmark test sets: 0.9076
• Human evaluation on 74 test-suite sentences: 53.67% correct

The 6 benchmark test sets used to compute the average BLEU and COMET values were:

• gold1
• gold2
• test-suite
• flores
• tatoeba
• taCoN

The first 3 datasets are our own, you can find the links here. Human evaluation was carried out by 3 native Galician annotators using binary correct/incorrect judgments.

These results support the main finding for es-gl: unlike more distant language pairs, Spanish-Galician reaches its best quality at an intermediate training size, while larger corpora can introduce source-language interference.

Intended use

This model is intended for:

• research on Spanish-Galician machine translation
• evaluation of data scaling in LLM fine-tuning
• practical translation workflows into Galician when using the Carvalho model family

Limitations

• This is a LoRA adapter, not a standalone merged model.
• It is specialized for Spanish-to-Galician translation and should not be treated as a general-purpose chat assistant.
• The model inherits the capabilities and limitations of the base model.
• As reported in the paper, translation quality depends strongly on language pair and corpus design; the best training size for es-gl should not be assumed to be optimal for other directions.