Naryn-1.0-7b

Naryn-1.0-7b is a high-performance, fine-tuned version of Mistral-7B-v0.3, optimized for general knowledge, logical reasoning, and sophisticated conversational capabilities. It was specifically developed to be lightweight enough to run on consumer-grade hardware.

🚀 Model Details

• Developed by: LazyLoopStudio
• Base Model: Mistral-7B-v0.3 (4-bit quantized via Unsloth)
• Architecture: Transformer with LoRA adapters
• Training Framework: Unsloth
• Quantization: GGUF (Q4_K_M) - Optimized for 8GB VRAM
• Context Length: 2048 tokens

🧠 Training Highlights

The model underwent fine-tuning using LoRA (Low-Rank Adaptation) with a focus on:

• Logic & Reasoning: Trained on the SlimOrca dataset to improve step-by-step thinking.
• General Knowledge: Enhanced factual accuracy for diverse topics.
• Eloquence: Refined conversational flow to provide human-like, helpful responses.

Training Hyperparameters:

• LoRA Rank (r): 32
• LoRA Alpha: 32
• Learning Rate: 1e-4
• Optimizer: Paged AdamW 8-bit
• Batch Size: 1 (with Gradient Accumulation Steps: 8)

📸 Vision Capabilities

While the base of Naryn-1.0-7b is a language model, it is designed to be used in Multimodal workflows. When paired with a compatible Vision Projector (like those used in LLaVA-Mistral), it can process and describe images effectively within environments like LM Studio or Ollama.

💻 Hardware Compatibility (RTX 4060)

This model is "4060-Native."

• Model Size: ~4.37 GB
• VRAM Usage (Idle): ~4.5 GB
• VRAM Usage (Peak): ~7.0 GB (including Vision overhead and KV-cache)
• Performance: Expect 40-60 tokens per second on an RTX 4060.

🛠 Usage

Via LM Studio

1. Search for LazyLoopStudio/Naryn-1-GGUF.
2. Download the Q4_K_M version.
3. In settings, ensure GPU Offload is set to "Max".

Via Ollama

Create a Modelfile with a wanted System Prompt, than run this command:

ollama create Naryn-1.0-7b -f Modelfile

⚠️ Disclaimer

Naryn-1.0-7b is a large language model. It can occasionally produce factually incorrect information (hallucinations). Always verify critical information.