mshojaei77/gemma-2-2b-fa-v2

Persian Gemma 2b (v2)

This repository hosts Persian Gemma 2b v2, an optimized version of the initial Persian Gemma 2b model. This version has undergone further experimental improvements through techniques like self-merging to enhance its performance and efficiency for Persian language conversational tasks.

This model builds upon the foundation of Google's Gemma-2-2b-it and the initial fine-tuning efforts of mshojaei77/Gemma-2b-fa.

Key Improvements in v2

• Optimized Performance: This version incorporates techniques to improve the model's performance in generating Persian text and engaging in conversations.
• Self-Merged: The model has been merged with itself, potentially leading to a more robust and coherent representation.

Important Note: This is still an experimental model and is under active development. While optimizations have been applied, it's crucial to understand that it retains the limitations inherent in a 2 billion parameter model and the early-stage nature of this project. Output quality may vary, and critical evaluation is still necessary.

How to Use

You can use this model just like the original mshojaei77/Gemma-2b-fa, using the transformers library and the pipeline.

import torch
from transformers import pipeline

# Initialize the text generation pipeline
pipe = pipeline(
    "text-generation",
    model="mshojaei77/gemma-2-2b-fa-v2",
    model_kwargs={"torch_dtype": torch.bfloat16},
    device="cuda",  # Or "mps" for Macs with Apple Silicon
)

# Prepare input messages (using the gemma chat template implicitly)
messages = [
    {"role": "user", "content": "سلام چطوری؟"},
]

# Generate a response with a maximum of 512 new tokens
outputs = pipe(messages, max_new_tokens=512, chat_template="gemma") # Explicitly using chat_template for clarity
assistant_response = outputs[0]["generated_text"][-1]["content"].strip()

print(assistant_response)
# Example Output (Illustrative - Output quality may vary significantly):
# سلام! من خوبم، ممنون. شما چطوری؟ 😊

Key Usage Notes:

• model="mshojaei77/gemma-2-2b-fa-v2": Make sure to specify the correct model name (v2) in your code.
• chat_template="gemma": Continue to use the gemma chat template for proper formatting.
• Experimental Nature: Expect variable output quality. This is still an experimental model.

Model Details (v2)

• Base Model: google/gemma-2-2b-it
• Previous Version: mshojaei77/Gemma-2b-fa
• Optimization Techniques: Self-Merging, Further Optimization.
• Architecture: Gemma2ForCausalLM (same as base model)
• Model Size: 2 billion parameters

Intended Use

This model is intended for:

• Research and Experimentation: Exploring the impact of Self-Merging techniques on Persian Gemma models.
• Educational Purposes: Demonstrating advanced fine-tuning and optimization methods.
• Community Development: Contributing to the growing ecosystem of Persian language models.
• Prototyping (with caution): For early-stage prototyping, acknowledging its experimental nature.

Limitations

While optimized, this model still has limitations:

• Experimental Stage: Under ongoing development and refinement.
• 2b Parameter Model: Performance is limited by the model size compared to larger models.
• Variable Output Quality: Output quality can still be inconsistent and require careful evaluation.
• Potential for Imperfections: May still exhibit issues like fluency problems, factual inaccuracies, or biases.

Use this model responsibly and be aware of its experimental nature.

Citation:

If you use this model in your research or applications, please cite it using the following DOI: 10.57967/hf/4772