ikergf/asturian-nllb-lora-subsampled-loss

asturian-nllb-lora-colab

LoRA adapter for Spanish→Asturian (spa_Latn → ast_Latn) machine translation, fine-tuned on top of NLLB-200-distilled-600M. This is the Colab prototyping model, trained on a subsampled dataset of 5,000 sentence pairs with loss-based validation only.

Training Data

• —Train: ES-AST Parallel Corpus (AINA) — 5,000 sentence pairs (subsampled)
• —Dev/Test: FLORES+ — 500 sentences each (subsampled)

Training Details

• —Base model: facebook/nllb-200-distilled-600M
• —Method: LoRA (r=8, alpha=16, target modules: q_proj, k_proj, v_proj, out_proj)
• —Batch size: 8
• —Learning rate: 5e-5
• —Epochs: 1.5
• —Validation: Loss only (no metric-based model selection)
• —Precision: float16

Results on FLORES+ devtest (subsampled, 500 sentences)

Note: This model underperforms the baseline (BLEU 15.32, chrF++ 42.17) due to the small training set.