🌐 Nyra-A: The Logic Core

Nyra-A is a specialized high-performance reasoning model developed by Logihertz Systems OPC Pvt Ltd. As part of the independent Nyra Project, this model serves as the "Primary Logic Core" (Tier A), optimized for mathematical consistency, structured data processing, and complex logical deduction.

🛠 Model Specifications

• Developer: Logihertz Systems
• Lead Architect: Sameer Tawade
• Project Status: Independent Research
• Architecture: Optimized Llama-3-8B (Transformer-based)
• Merge Methodology: DARE-TIES + SLERP (Optimized for weight-sum stability)
• Language(s): English (Primary)

🎯 Intended Use Cases

Nyra-A is engineered for standalone applications requiring high precision:

• Algorithmic Reasoning: Solving complex mathematical and logical proofs.
• Structured Output: Generating precise JSON, XML, and complex code structures.
• Analytical Processing: Acting as a refiner for complex multi-turn instructions where hallucination must be minimized.

📊 Evaluation & Benchmarking Matrix

This model is currently undergoing rigorous evaluation. Scores are marked as pending while the self-verified evaluation pipeline completes.

Category	Benchmark	Metric	Score	Status
General Reasoning	MMLU-Pro	5-shot Accuracy	Pending	Eval in Progress
Math Execution	GSM8K	8-shot Strict Match	Pending	Eval in Progress
Advanced Math	MATH	4-shot Chain-of-Thought	Pending	Eval in Progress
Graduate Logic	GPQA	0-shot Accuracy	Pending	Eval in Progress
Code Reasoning	HumanEval	Pass@1	Pending	Eval in Progress

💻 Implementation

To run Nyra-A locally, ensure you have the latest transformers library installed.

from transformers import AutoModelForCausalGeneration, AutoTokenizer
import torch

model_id = "logihertz/nyra-A"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

prompt = "Analyze the efficiency of a recursive function versus an iterative approach."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

⚖️ Limitations & Ethical Considerations

Nyra-A is released under the Llama 3 Community License. While heavily optimized for logic, it may still exhibit occasional hallucinations or inherit biases from its foundational weights. Users should implement secondary validation systems for critical, public-facing deployments.