This model is a fine-tuned version of meta-llama/Llama-3.2-3B using the GRIT (Geometric Reprojection Instruction Tuning) algorithm and LoRA on the google/boolq dataset.

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

🚀 Training Details

GRIT Algorithm

K-FAC Updates: Every 10 steps (adaptive) for second-order preconditioning.
Neural Reprojection: Every 20 steps (adaptive) for rank optimization.
Rank Adaptation: Enabled (Threshold: 0.9, Min Rank: 4).
Optimized LoRA Modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']

Fine-tuning Configuration

Base Model: meta-llama/Llama-3.2-3B
Quantization: 4-bit (NF4) with bf16 compute.
LoRA Rank: 16
LoRA Alpha: 32
Batch Size: 8 (per device)
Gradient Accumulation: 4 (Effective batch = 32)
Learning Rate: 2.0e-05
Precision: bf16 mixed precision
Sequence Length: 1024 tokens
Gradient Checkpointing: Enabled

Performance Improvements

✅ Faster Convergence: K-FAC preconditioning aligns updates with curvature.
✅ Adaptive Rank: Dynamically prunes LoRA rank to improve parameter efficiency.

📊 Training Metrics

Total Steps: 295
Final Loss: 0.3212439351162668
Trainable Params: 24,313,856

📝 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. The use of Unsloth further accelerates this process.

📝 Citation

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

bibtex

@misc{grit-lora-Llama-3.2-3B-boolq},
  title={ meta-llama/Llama-3.2-3B Fine-tuned with GRIT on google/boolq },
  author={D1zzYzz},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/D1zzYzz/GRIT-BOOLQ-QLORA-llama-3.2-3B-Energy-0.9}
}

⚖️ License

This model inherits the Apache 2.0 license.