---
language:
- dv
- en
pipeline_tag: translation
library_name: transformers
license: apache-2.0
base_model: google/mt5-large
tags:
- dhivehi
- thaana
- maldives
- translation
metrics:
- chrf
---

# mT5-large English→Dhivehi (sentence-level)

mT5-large (1.2B, encoder-decoder) for **English→Dhivehi** translation, trained on a sentence-level Dhivehi–English corpus (machine-translated).

Sentence-level variant. For multi-sentence input, see the paragraph models: [Qwen-para](https://huggingface.co/Neobe/en-dhivehi-qwen3-4b-lora-paragraph) / [mT5-para](https://huggingface.co/Neobe/en-dhivehi-mt5-large-paragraph).
Suited for single-sentence input.

## Scores (chrF / chrF++ / BLEU)
| Benchmark | chrF | chrF++ | BLEU |
|---|---|---|---|
| **gold** (human references, article-level, N=500) | 50.47 | 42.17 | 4.46 |
| held-out chunk (in-distribution) | 71.97 | 64.85 | 24.77 |
| held-out sentence (in-distribution) | 68.72 | 61.82 | 22.96 |

chrF is the metric to trust for Thaana; BLEU is unreliable (word segmentation / morphology).

## Example
> **Input (en):** While this is a 7.6 percent increase compared to the same period last year, an average of 7,778 tourists visit the Maldives daily.
>
> **Output (dv):** މިއީ މިދިޔަ އަހަރުގެ މި މުއްދަތާ އަޅާބަލާއިރު 7.6 އިންސައްތައިގެ ކުރިއެރުމެއް ކަމަށްވާއިރު، ދުވާލަކު އެވްރެޖްކޮށް 7،778 ފަތުރުވެރިން ރާއްޖެ ޒިޔާރަތްކުރެއެވެ.

_Real held-out sample and this model's own output._

## Usage
```python
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

m = "Neobe/en-dhivehi-mt5-large-sentence"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSeq2SeqLM.from_pretrained(m, torch_dtype=torch.float32).eval().cuda()  # fp32

src = "The President of the Maldives met with the cabinet today."
inp = tok(src, return_tensors="pt", truncation=True, max_length=1024).to("cuda")
out = model.generate(**inp, max_new_tokens=512, num_beams=4, length_penalty=1.0, no_repeat_ngram_size=0)
print(tok.decode(out[0], skip_special_tokens=True))
```

> **Run in fp32** — the T5 family produces garbage in bf16/fp16. ~5–6 GB VRAM.

## Training
Base `google/mt5-large`; fp32; Adafactor; LR 1e-4 cosine; max_length 512; 1 epoch; effective batch ~32; gradient checkpointing.

## Limitations
Domain = Maldivian news / press / Wikipedia; technical or informal English is out of distribution. Non-human references are machine-generated (distillation).

## Citation
```bibtex
@misc{neobe_en_dhivehi_mt5_large_sentence_2026,
  title  = {mT5-large English→Dhivehi (sentence-level)},
  author = {Neobe},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/Neobe/en-dhivehi-mt5-large-sentence}}
}
```