README.md

---
license: unknown
language:
- si
metrics:
- perplexity
library_name: transformers
tags:
- AshenBerto
- Sinhala
- Roberta
---



### 🌟 Overview  

This is a slightly smaller model trained on half of the [FastText](https://fasttext.cc/docs/en/crawl-vectors.html) dataset. Since Sinhala is a low-resource language, there’s a noticeable lack of pre-trained models available for it. 😕 This gap makes it harder to represent the language properly in the world of NLP.  

But hey, that’s where this model comes in! 🚀 It opens up exciting opportunities to improve tasks like sentiment analysis, machine translation, named entity recognition, or even question answering—tailored just for Sinhala. 🇱🇰✨  

---

### 🛠 Model Specs  

Here’s what powers this model (we went with [RoBERTa](https://arxiv.org/abs/1907.11692)):  

1️⃣ **vocab_size** = 25,000  
2️⃣ **max_position_embeddings** = 514  
3️⃣ **num_attention_heads** = 12  
4️⃣ **num_hidden_layers** = 6  
5️⃣ **type_vocab_size** = 1  
🎯 **Perplexity Value**: 3.5  

---

### 🚀 How to Use  

You can jump right in and use this model for masked language modeling! 🧩  

```python
from transformers import AutoTokenizer, AutoModelWithLMHead, pipeline

# Load the model and tokenizer
model = AutoModelWithLMHead.from_pretrained("ashenR/AshenBERTo")
tokenizer = AutoTokenizer.from_pretrained("ashenR/AshenBERTo")

# Create a fill-mask pipeline
fill_mask = pipeline('fill-mask', model=model, tokenizer=tokenizer)

# Try it out with a Sinhala sentence! 🇱🇰
fill_mask("මම ගෙදර <mask>.")
```