Pritish92/open-llama-3b-v2-grit-alpaca

OpenLlama-3B-v2 Fine-tuned with GRIT and QLoRA

This model is a fine-tuned version of openlm-research/open_llama_3b_v2 using the GRIT (Gradient Regularized Instruction Tuning) algorithm and QLoRA on the Alpaca dataset.

The base model is quantized to 4-bit (NF4) to enable efficient fine-tuning.

🚀 Training Details

GRIT Algorithm

• K-FAC Updates: Every 200 steps for second-order preconditioning
• Neural Reprojection: Every 500 steps for rank optimization
• Optimized LoRA Modules: attention + key MLP layers (as per design)

Fine-tuning Configuration

• Base Model: OpenLlama 3B v2
• Quantization: 4-bit (NF4) with float16 compute
• LoRA Rank: 64
• LoRA Alpha: 128
• Batch Size: 16 (per device)
• Gradient Accumulation: 4 (Effective batch = 64)
• Learning Rate: 5.0e-05
• Precision: bf16 mixed precision
• Sequence Length: 512 tokens
• Gradient Checkpointing: Enabled

Performance Improvements

• ✅ Faster Convergence: K-FAC preconditioning aligns updates with curvature
• ✅ Memory-Efficient: 4-bit quantization (QLoRA) and gradient checkpointing used.
• ✅ Efficient Training: Utilizes accelerate for efficient training.

📊 Training Metrics

• Total Steps: 732
• Final Training Loss: 0.2282
• Final Validation Loss: 0.22849
• BLEU (val): 0.2452
• Trainable Params: 42,598,400 (1.23% of total)

🏷️ Model Tags

• Instruction-tuned with GRIT and QLoRA
• GRIT-tuned Model
• 4-bit Quantized Model
• LoRA rank 64
• Mixed precision (bf16)
• Alpaca dataset fine-tuning

📝 Algorithm Details

• K-FAC Preconditioning (Natural Gradient) and Neural Reprojection as per GRIT method
• Memory Efficient: Covariance matrices on CPU to reduce GPU load

🏆 Results

In benchmark comparisons, GRIT has shown faster convergence and better stability than standard LoRA or fine-tuning, making it well-suited for efficient single-epoch training.

📝 Citation

If you use this model, please cite:

@misc{grit-openllama-3b-alpaca,
  title={OpenLlama 3B v2 Fine-tuned with GRIT on Alpaca},
  author={Pritish92},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/Pritish92/open-llama-3b-v2-grit-alpaca}
}

⚖️ License

This model inherits the Apache 2.0 license.