A LoRA adapter for Tamil to Sinhala machine translation, trained on the TamSiPara Tamil-Sinhala parallel corpus (Pushpananda et al., 2024) as part of the SinGen Sinhala text generation benchmark.

The adapter is trained on top of google/gemma-4-31B-it using the model's own chat template, with the same instruction wording as the SinGen zero-shot and few-shot prompting runs.

Usage

python

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel

processor = AutoProcessor.from_pretrained("google/gemma-4-31B-it")
model = AutoModelForImageTextToText.from_pretrained("google/gemma-4-31B-it", dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "sinhala-nlp/gemma-4-31B-it-TamSiPara-Ta2Si-en")

(Use AutoTokenizer / AutoModelForCausalLM for text-only checkpoints.)

The model continues from the Translation: prefix inside the assistant turn.

Training


Training pairs	21624
Instruction language	`en`
Epochs	3.0
Effective batch size	16
Learning rate	0.0002
Max sequence length	768
LoRA r / alpha / dropout	16 / 32 / 0.05
Target modules	language-model linear layers (vision tower excluded)

Evaluation

Held-out tail of TamSiPara (999 sentence pairs), whitespace-tokenized (sacreBLEU's default 13a tokenizer splits Sinhala conjuncts and vowel signs):

Metric	Score
Corpus sacreBLEU	17.91
Sentence-level BLEU mean	9.74

Licence

Derived from Gemma and governed by the Gemma Terms of Use.