Neura-1

Neura 1 - Next-Generation AI Assistant

Neura 1 represents a significant leap in AI capability, featuring extended training, superior intelligence, and cutting-edge awareness.

🌟 Key Features

• Current Data Awareness: Trained with data up to 2025-10-31
• Optimized Training: 90 steps for balanced performance and efficiency
• Maximum Capacity: LoRA r=256, alpha=512 for exceptional learning
• Premium Data: 2,500+ samples from 5 elite public datasets
• Advanced Reasoning: Enhanced mathematical and logical capabilities
• Extended Context: 8,192 token sequence length

📊 Training Specifications

Dataset Composition (2,600 samples)

• OpenOrca: 600 samples (instruction following)
• Ultrafeedback: 600 samples (quality responses)
• WizardLM Evol: 500 samples (complex reasoning)
• No Robots: 500 samples (diverse conversations)
• MetaMathQA: 400 samples (mathematical reasoning)

Training Configuration

• Base Model: Llama 3.1 8B Instruct
• Training Steps: 90 (optimized for efficiency)
• Learning Rate: 5e-5 (optimized)
• Effective Batch Size: 16
• LoRA Rank: 256 (maximum capacity)
• LoRA Alpha: 512 (maximum adaptation)
• Sequence Length: 8,192 tokens
• Training Time: ~8.9 minutes
• Final Loss: 1.2700

🚀 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("arthu1/Neura-1")
tokenizer = AutoTokenizer.from_pretrained("arthu1/Neura-1")

messages = [
    {"role": "system", "content": "You are Neura 1, a cutting-edge AI assistant."},
    {"role": "user", "content": "Your question here"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=384, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

💡 Capabilities

• Advanced mathematical reasoning
• Complex logical analysis
• Superior code generation
• In-depth explanations
• Current event awareness
• Multi-step problem solving

⚙️ Optimization Details

• Optimizer: AdamW 8-bit
• Scheduler: Cosine with warmup
• Warmup Steps: 10
• Weight Decay: 0.01
• Max Gradient Norm: 0.5
• Gradient Checkpointing: Enabled
• Mixed Precision: BF16/FP16

📅 Training Information

• Training Date: 2025-10-31
• Data Cutoff: 2025
• Total Parameters: ~8B (4-bit quantized)
• Trainable Parameters: ~134M (LoRA)

🎯 Performance

Expected performance characteristics:

• Superior reasoning capabilities
• Enhanced mathematical accuracy
• Improved code generation
• Better context understanding
• Stable and consistent responses

---

Neura 1 - Where Intelligence Meets Stability