Aura-1-Lightning

Aura-1-Lightning ⚡

A high-performance, structurally distilled 7B causal language model optimized for rapid local inference and reasoning density. Developed under the Waveforce AI framework, this model was trained using a custom hybrid CPU/GPU cross-device execution pipeline designed to maximize optimization stability under tight hardware resource thresholds.

---

📖 Model Card

🔬 Architectural Strategy & Distillation Profile

Traditional quantization often degrades a model's latent logic representation. Aura-1-Lightning avoids this degradation by employing True Cross-Device Probabilistic Logit Alignment.

During training, the student model's base parameters were frozen utilizing a dual 4-bit QLoRA configuration to fit execution within standard hardware limits. Concurrently, the unreduced logit streams from both the student and teacher models were cross-compared natively on their respective devices. By slicing matching vocabulary spaces and routing the heavy matrix operations directly to CPU memory, the training loop successfully optimized structural Kullback-Leibler (KL) divergence alongside hard cross-entropy targets without memory overflow.

Model Ancestry & Credits

To give credit to the engineering teams behind the base architectures:

• Student Base Model: Qwen/Qwen2.5-7B-Instruct — Celebrated for its massive vocabulary size ($151,643$ tokens) and base coding and mathematical reasoning density.
• Teacher Foundation Model: google/gemma-2-9b-it — Providing the high-precision probabilistic curves that guided the student's hidden states.

---

📈 Performance & Hardware Footprint

By compiling the learned adapter weights directly back into the unquantized base layers using full precision math, the final repository is delivered as a pure standalone model (native .safetensors format). It requires no custom wrappers or loose adapter configurations to run.

VRAM Allocation Metrics

• Base Weight Footprint (FP16/BF16): ~14.5 GB (Fits within single-card consumer GPUs, including RTX 3090/4090, T4, or L4 systems).
• Quantized Deployment (GGUF / Int4 / NF4): ~4.5 GB to 5.2 GB (Optimized for low latency execution on local laptops, embedded devices, and local AI workflows like AnythingLLM).
• Context Optimization: Fully configured with native .jinja chat templates supporting strict multi-turn dialogue boundaries up to its optimized token windows.

---

📦 Repository Structure

The aura-1-lightning-weights directory contains all components required for native execution:

waveforce-ai/Aura-1-Lightning/
└── aura-1-lightning-weights/
    ├── model.safetensors         # Compiled, unquantized standalone weights
    ├── config.json               # Native layer and attention head geometry
    ├── generation_config.json    # Default inference sampling hyper-parameters
    ├── tokenizer.json            # Vocabulary token mappings (151k token capacity)
    ├── tokenizer_config.json     # Chat layout definitions
    └── chat_template.jinja       # Structural instruction template wrapper