This repository contains a version of Google's Gemma-3 270M IT model, fine-tuned for conversational instruction-following. The fine-tuning was performed using Unsloth for significantly faster training and reduced memory usage, making it possible to train on consumer-grade hardware.

The model was trained on the nvidia/HelpSteer2 dataset to improve its ability to generate helpful, safe, and contextually relevant responses in interactive chat scenarios.

⚙️ Model Details

Base Model: unsloth/gemma-3-270m-it
Fine-tuning Method: LoRA adaptation via PEFT
Dataset: nvidia/HelpSteer2
Frameworks:
- Unsloth
- Hugging Face Transformers
- TRL

🚀 How to Use

You can easily run this model using the transformers library. Make sure you have transformers, unsloth, and torch installed.

python

from transformers import pipeline
import torch

# Load the model using the text-generation pipeline
# Make sure to use your Hugging Face username and repo name
generator = pipeline(
    "text-generation",
    model="farhanahmedconnect/gemma-3_finetuned",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Use the Gemma-3 chat template
messages = [
    {"role": "user", "content": "What are the top 3 benefits of using Unsloth for model training?"},
]

# Generate a response
outputs = generator(
    messages,
    max_new_tokens=256,
)

# Print the generated text from the assistant
print(outputs[0]["generated_text"][-1]['content'])

🔧 Training Details

Training Data

The model was fine-tuned using the high-quality conversational dataset nvidia/HelpSteer2, which is specifically designed for instruction-following tasks. The dataset includes dialogues formatted into clear user and assistant turns, allowing the model to learn how to generate helpful, safe, and contextually relevant responses.

The training process utilized a masking technique, ensuring that the loss calculation focused solely on the assistant's replies, making the training more efficient and better aligned with conversational objectives.

Training Procedure

The fine-tuning was performed using Unsloth's FastLanguageModel and the SFTTrainer from the TRL library, which optimize both training speed and memory usage. The process involved several key steps:

Model Loading: The base model was loaded and optimized using unsloth.FastModel to ensure faster performance and memory efficiency.
LoRA Adapters: LoRA (Low-Rank Adaptation) adapters were applied for parameter-efficient fine-tuning, allowing for faster adaptation of the model with fewer parameters.
Training: The model was trained using the SFTTrainer, which enabled supervised fine-tuning specifically focused on instruction-following.
Final Model Saving: The fine-tuned LoRA adapters were saved as separate artifacts, which can easily be merged with the base model for deployment, ensuring seamless model updates and scalability.

Final Artifacts

The fine-tuned LoRA adapters and model checkpoint are available for deployment and further fine-tuning, ensuring adaptability to evolving use cases and datasets.
These artifacts provide a high-quality, efficient model that balances speed, memory efficiency, and conversational accuracy.

📊 Performance

This model demonstrates improved instruction-following capabilities compared to the base model, with enhanced performance in:

Conversational dialogue generation
Context-aware responses
Safety and helpfulness alignment
Reduced hallucinations in chat scenarios

🛠️ Training Configuration

yaml

training_details:
  base_model: unsloth/gemma-3-270m-it
  dataset: nvidia/HelpSteer2
  method: LoRA (Low-Rank Adaptation)
  framework: 
    - Unsloth FastLanguageModel
    - TRL SFTTrainer
  optimization:
    - Parameter-efficient fine-tuning
    - Memory usage optimization
    - Training speed enhancement
  masking: Assistant replies only
  artifacts: LoRA adapters + model checkpoint

📄 License

This model is released under the Apache 2.0 license, following the licensing terms of the base Gemma model.

🙏 Acknowledgments

Google for the Gemma-3 base model
Unsloth for the efficient training framework
NVIDIA for the HelpSteer2 dataset
Hugging Face for the transformers library and model hosting