NRS NEURAL REASONING SYSTEM

SKT AI LABS

REQUEST

🛠️ How to Use & Run

This model is optimized for both Local Execution and Cloud Notebooks (Colab).

1. 🐍 Python (Transformers)

Install the required libraries:

pip install transformers torch accelerate

Basic Inference Code:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "SKT-NRS/NRS_QWEN_MYTHOS_1M" 

# Load Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Prepare Prompt
messages = [
    {"role": "system", "content": "You are NRS, an advanced reasoning assistant."},
    {"role": "user", "content": "Explain quantum entanglement simply."}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate
outputs = model.generate(
    **inputs,
    max_new_tokens=4096,
    temperature=0.6,
    top_p=0.95,
    do_sample=True
)

response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)
print(response)

2. ☁️ Google Colab Ready

Run this on a free T4 GPU or paid A100/V100 instances.

!pip install transformers accelerate bitsandbytes

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "SKT-NRS/NRS_QWEN_MYTHOS_1M"

# Load in 4-bit for memory efficiency
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    load_in_4bit=True,
    device_map="auto",
    torch_dtype=torch.float16
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

print("Model Loaded Successfully! Ready for Reasoning.")

3. 🖥️ Local Running (Ollama / LM Studio)

For the best local experience, use GGUF quantizations.

Using Ollama:

ollama create nrs-mythos -f Modelfile
ollama run nrs-mythos

Using LM Studio:

1. Download the .gguf file from the "Files" section.
2. Drag and drop into LM Studio.
3. Set Context Length to 8192 or higher.

4. ⚡ High-Performance Serving (vLLM)

For production-grade speed and 1M context support:

pip install vllm

vllm serve SKT-NRS/NRS_QWEN_MYTHOS_1M \
    --max-model-len 1000000 \
    --gpu-memory-utilization 0.9 \
    --dtype bfloat16

🧠 Technical Details & Training

Base Model

• Architecture: Qwen 3.5 9B
• Context Window: 1,048,576 tokens (via YaRN RoPE Scaling)

NRS Enhancement Process

The model underwent a rigorous Neural Reasoning System (NRS) enhancement pipeline:

1. Reasoning Boosting Tool: Proprietary NRS tools generated high-quality Chain-of-Thought (CoT) data.
2. Supervised Fine-Tuning (SFT): Tuned on ~500k high-quality reasoning samples (coding, math, logic).
3. Tool Calling Optimization: Enhanced native function calling for Python & Web Search.

Sampling Parameters

• Temperature: 0.6
• Top_P: 0.95
• Top_K: 20
• Repetition Penalty: 1.05

⚠️ Limitations & Disclaimer

• Reasoning Mode: The model outputs <think> blocks. Parse them if needed.
• Uncensored Nature: Designed for open research. Use responsibly.
• Hallucinations: Always verify critical facts with external sources.

LICENSE AND TERMS