README.md

---
license: mit
task_categories:
- text2text-generation
language:
- ml
size_categories:
- 1K<n<10K
---

# Romanized to Native Malayalam Script Transliteration

Shared-Task code submitted to [IndoNLP Workshop](https://indonlp-workshop.github.io/IndoNLP-Workshop/) colocated with [COLING 2025](https://coling2025.org/), in Abu Dhabi, UAE.

The training notebook shows the [code](https://github.com/VRCLC-DUK/ml-en-transliteration).

## Training

Data derived from:

1. Dakshina
    1. [Romanized](https://huggingface.co/datasets/vrclc/Dakshina-romanized-ml) - 186 thousand  word transliteration pairs
    2. [Lexicons] (https://huggingface.co/datasets/vrclc/dakshina-lexicons-ml) - 58.4 thousand lexicon entries
2. [Aksharantar](https://huggingface.co/datasets/vrclc/Aksharantar-ml) - 4.3 million word transliteration pairs


## RESULT

We present the evaluation scores averaged over the two test datsets shared by the task organizers.


| **Dataset**   | **CER (%)** | **WER (%)** | **BLEU (%)** |
|---------------|-------------|-------------|--------------|
| Test Set-1    | 7.4         | 34.5        | 32.7         |
| Test Set-2    | 22.7        | 66.9        | 7.5          |

### Test Set-1


![Distribition of Evaluation Scores on Test Set-1](test-1.png "Distribition of Evaluation Scores on Test Set-1")

### Test Set-2


![Distribition of Evaluation Scores on Test Set-2](test-2.png "Distribition of Evaluation Scores on Test Set-2")