91 lines
2.1 KiB
Markdown
91 lines
2.1 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- fa
|
|||
|
|
tags:
|
|||
|
|
- translation
|
|||
|
|
- english-to-persian
|
|||
|
|
- persian-to-english
|
|||
|
|
- bilingual
|
|||
|
|
- farsi
|
|||
|
|
- persian
|
|||
|
|
model_type: llama
|
|||
|
|
base_model: meta-llama/Llama-3.2-1B
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library: transformers
|
|||
|
|
widget:
|
|||
|
|
- text: |
|
|||
|
|
### English:
|
|||
|
|
The children were playing in the park.
|
|||
|
|
### Persian:
|
|||
|
|
- text: |
|
|||
|
|
### Persian:
|
|||
|
|
من به مدرسه میروم.
|
|||
|
|
### English:
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
[](https://huggingface.co/spaces/Sheikhaei/llama-3.2-1b-english-persian-translator)
|
|||
|
|
|
|||
|
|
|
|||
|
|
# LLaMA 3.2 1B – English ↔ Persian Translator
|
|||
|
|
|
|||
|
|
This model is a fine-tuned version of [`meta-llama/Llama-3.2-1B`](https://huggingface.co/meta-llama/Llama-3.2-1B), trained for **bidirectional translation** between **English and Persian**. It supports both:
|
|||
|
|
|
|||
|
|
- 🇬🇧 English → 🇮🇷 Persian
|
|||
|
|
- 🇮🇷 Persian → 🇬🇧 English
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Format
|
|||
|
|
|
|||
|
|
The model expects prompts in the following format:
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
### English:
|
|||
|
|
The children were playing in the park.
|
|||
|
|
### Persian:
|
|||
|
|
```
|
|||
|
|
or
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
### Persian:
|
|||
|
|
کودکان در پارک بازی میکردند.
|
|||
|
|
### English:
|
|||
|
|
```
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Usage
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained("Sheikhaei/llama-3.2-1b-en-fa-translator", torch_dtype="auto", device_map="auto")
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained("Sheikhaei/llama-3.2-1b-en-fa-translator")
|
|||
|
|
|
|||
|
|
prompt = """### English:
|
|||
|
|
The children were playing in the park.
|
|||
|
|
### Persian:
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
|||
|
|
outputs = model.generate(**inputs, max_new_tokens=100, do_sample=False)
|
|||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Training Data
|
|||
|
|
|
|||
|
|
This model was fine-tuned on a custom English–Persian parallel dataset containing ~640,000 sentence pairs. The source data was collected from Tatoeba and then translated and expanded using the Gemma-3-12B model.
|
|||
|
|
|
|||
|
|
## Evaluation
|
|||
|
|
|
|||
|
|
| Direction | BLEU | COMET |
|
|||
|
|
|---------------|------|------|
|
|||
|
|
| English → Persian | 0.47 | 0.89 |
|
|||
|
|
| Persian → English | 0.58 | 0.91 |
|
|||
|
|
|
|||
|
|
## License
|
|||
|
|
|
|||
|
|
Apache 2.0
|
|||
|
|
|