You can deploy this model and start earning money today!

mdabucse/
Gemma-3b-Tanglish-Model

Name: Gemma-3b-Tanglish-Model
Availability: InStock

safetensors

gemma3_text

Views

No views yet

🧠 Gemma Fine-Tuned Model — Tamil/English → Tanglish Conversion

📘 Model Overview

This model is a fine-tuned version of Google's Gemma model, trained for Tamil and English to Tanglish (Romanized Tamil) conversion.
It follows an instruction-based fine-tuning format inspired by Alpaca/Vicuna-style instruction datasets.

🧩 Task

Primary Objective: Convert mixed Tamil/English text into Tanglish (Romanized Tamil).
Example:
- Input: "நீங்கள் எப்படி இருக்கிறீர்கள்?"
- Output: "Neenga eppadi irukeenga?"

⚙️ Installation & Setup

pip install torch transformers accelerate

🚀 Load Model from Hugging Face

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "mdabucse/gemma-finetuned-tanglish"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

💬 Input Injection Template

This is the prompt structure used for both training and inference:

python

test_text = """### Instruction:
Convert Tamil/English to Tanglish (Romanized Tamil).
### Input:
{context}
### Response:"""

You can inject your own input dynamically:

python

context = "நீங்கள் எப்படி இருக்கிறீர்கள்?"
prompt = test_text.format(context=context)

🧠 Example Inference

python

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    temperature=0.7,
    top_p=0.9,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧾 Sample Output

### Instruction:
Convert Tamil/English to Tanglish (Romanized Tamil).
### Input:
நீங்கள் எப்படி இருக்கிறீர்கள்?
### Response:
Neenga eppadi irukeenga?

🧰 Parameters Explanation

Parameter	Description
`max_new_tokens`	Maximum number of tokens to generate.
`temperature`	Controls randomness (higher = more creative).
`top_p`	Nucleus sampling cutoff for probabilistic control.
`device_map`	Automatically maps model to available devices (GPU/CPU).
`torch_dtype`	Uses optimal data type (e.g. float16) for faster inference.

🧑‍💻 Training Notes

Base Model: google/gemma-2b or google/gemma-7b
Objective: Instruction fine-tuning for transliteration tasks.
Data Type: JSON/CSV with fields: instruction, input, output
Tokenizer: Gemma tokenizer with added Tamil tokens if needed.

📄 License

This model inherits the Gemma license and the dataset license used for fine-tuning.
Check respective Hugging Face page for redistribution rights.

🏷️ Metadata

Author: Mohamed Abubakkar S
Framework: 🤗 Unsloth
Task Category: Text-to-Text (Transliteration / Romanization)

mdabucse/Gemma-3b-Tanglish-Model

🧠 Gemma Fine-Tuned Model — Tamil/English → Tanglish Conversion

📘 Model Overview

🧩 Task

⚙️ Installation & Setup

🚀 Load Model from Hugging Face

💬 Input Injection Template

🧠 Example Inference

🧾 Sample Output

🧰 Parameters Explanation

🧑‍💻 Training Notes

📄 License

🏷️ Metadata

mdabucse/
Gemma-3b-Tanglish-Model