---
language:
- si
- pi
base_model: Qwen/Qwen3.5-9B
library_name: peft
tags:
- sinhala
- pali
- translation
- machine-translation
- singen
- lora
datasets:
- sinhala-nlp/pali-sinhala
metrics:
- bleu
---

# Qwen3.5-9B-PaliSinhala-Pali2Si-si

A LoRA adapter for **Pali to Sinhala translation**, fine-tuned from
[Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) as part of the SinGen Sinhala text
generation benchmark.

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "sinhala-nlp/Qwen3.5-9B-PaliSinhala-Pali2Si-si")
```

Prompts use the base model's chat template with thinking disabled, and the assistant
response begins with the `Translation:` prefix.

## Training

| | |
|---|---|
| Training pairs | 9450 |
| Instruction language | `si` |
| Length-matched subsample | False |
| Epochs | 1.0 |
| Effective batch size | 16 |
| Learning rate | 0.0002 |
| Max sequence length | 3072 |
| LoRA r / alpha / dropout | 16 / 32 / 0.05 |
| Thinking during training | False |

## Evaluation

Last 1000 rows of `sinhala-nlp/pali-sinhala`, whitespace-tokenized (sacreBLEU's default
13a tokenizer splits Sinhala conjuncts and vowel signs):

| Metric | Score |
|---|---|
| Corpus sacreBLEU | 0.00 |
| Sentence-level BLEU mean | 0.07 |
| Hyp/ref length ratio | 0.028 |

**Read the scores with the length ratio.** The corpus is in canonical order, so the
trailing 1000 rows used as the test set are much longer than the training pool
(median target length differs by roughly an order of magnitude). Predictions are
therefore shorter than references and BLEU is partly driven by the brevity penalty
rather than translation quality. This split is kept as-is so the numbers stay
comparable across the model families evaluated in SinGen.

## Licence

This adapter inherits the licence of the base model; check the base model card
before redistributing. Verify the terms of the `sinhala-nlp/pali-sinhala` dataset
on its dataset card as well.