---
license: apache-2.0
datasets:
- 24-mohamedyehia/gloss2text-Ar-sft
language:
- ar
base_model:
- google/gemma-3-270m-it
pipeline_tag: text-generation
tags:
- sign-language
- arabic
- ArSL
- gloss-to-text
- gemma-3
- accessibility
---

# Gloss2Text-V1-Gemma3-270M (Merged)

## Overview
**Gloss2Text-V1-Gemma3-270M** is a fine-tuned version of Google's **Gemma-3-270m-it**, optimized for **Arabic Sign Language (ArSL) gloss-to-text translation**.

This release is **merged**, which means the LoRA adapters have already been fused into the base model weights for faster inference and easier plug-and-play usage.

## Model Description
- **Task:** Converts a sequence of Arabic glosses into a natural, grammatically correct Modern Standard Arabic (MSA) sentence.
- **Input format:** Arabic gloss text, for example: "أنا شرب ماء الآن"
- **Output format:** Clean MSA text, for example: "أنا أشرب الماء الآن."
- **Architecture:** Gemma-3, 270M parameters
- **Training method:** Supervised fine-tuning with LoRA (rank 64), then merged

## Training Highlights
The model was trained on a specialized dataset containing diverse ArSL gloss-sentence pairs.
- **Final eval loss:** ~0.34
- **Precision:** Trained with `bf16` for improved numerical stability


## How to Use
Because this is a merged model, you can load it directly with the `transformers` library without needing `peft`:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "24-mohamedyehia/Gloss2Text-V1-Gemma3-270M"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

def translate_gloss(gloss_text):
    prompt = f"Translate ArSL gloss to an MSA sentence.\nGloss: {gloss_text}\nOutput: "
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=50)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)


# Example
print(translate_gloss("أنا ذهاب صيدلية"))
```


## Developer

Developed by **Mohamed Yehia**.

- **LinkedIn:** [Mohamed Yehia](https://www.linkedin.com/in/24-mohamed-yehia/)