gemma-3-1b-it-4bit-lora-dpo-aligned

This model is a fine-tuned version of google/gemma-3-1b-it using Direct Preference Optimization (DPO) on the ultrafeedback_binarized dataset.

Training Details

• Base Model: google/gemma-3-1b-it
• Fine-tuning Method: DPO (Direct Preference Optimization)
• Dataset: HuggingFaceH4/ultrafeedback_binarized
• Training Samples: 1000
• Evaluation Samples: 100
• Epochs: 1
• Batch Size: 1 (per device)
• Gradient Accumulation: 4
• Learning Rate: 5e-5
• Beta (DPO): 0.1
• Max Length: 1024
• Optimizer: adamw_8bit
• Precision: bfloat16
• Quantization: 4-bit NF4 (during training)
• LoRA Rank: 16
• LoRA Alpha: 32
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Memory Optimizations

• Gradient Checkpointing
• 8-bit AdamW Optimizer
• Pre-computed Reference Log Probabilities
• LoRA Parameter-Efficient Fine-tuning

Intended Use

This model is intended for text generation tasks with improved alignment through DPO training. It maintains the capabilities of the base Gemma 3 model while being better aligned with human preferences.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "gemma-3-1b-it-4bit-lora-dpo-aligned"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Example inference
prompt = "Explain quantum computing in simple terms."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Limitations

• This model inherits limitations from the base Gemma 3 model
• DPO alignment may not cover all edge cases or preferences
• Performance may vary on different hardware configurations

Citation

If you use this model, please cite the original Gemma model and the DPO paper:

@misc{gemma3,
  title={Gemma 3},
  author={Google DeepMind},
  year={2026}
}

@article{rafailov2023direct,
  title={Direct Preference Optimization: Your Language Model is Secretly a Reward Model},
  author={Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Manning, Christopher D and Finn, Chelsea and Ermon, Stefano},
  journal={arXiv preprint arXiv:2305.18290},
  year={2023}
}