Overview of Wraith-8B

Wraith-8B, developed by VANTA Research, is the inaugural model in their "Entity Series," a specialized fine-tune of Meta's Llama 3.1 8B Instruct. This 8 billion parameter model is engineered to excel in mathematical reasoning and STEM analysis, demonstrating a significant +37% relative improvement over its base model on the GSM8K benchmark, achieving 70% accuracy. It also shows enhanced factual accuracy with 58.5% on TruthfulQA and strong performance in MMLU Social Sciences (76.7%).

Key Capabilities

Superior Mathematical Reasoning: Achieves 70% on GSM8K, making it highly effective for arithmetic, algebra, and physics calculations.
Distinctive Personality: Features a "cosmic intelligence" persona, designed to enhance rather than hinder its analytical capabilities.
Enhanced Factual Accuracy: Improved TruthfulQA scores indicate better grounding and reduced hallucination.
Optimized Inference: Available in production-ready GGUF quantizations (e.g., Q4_K_M) for efficient deployment on consumer hardware.
Broad Context Window: Supports a context length of 131,072 tokens, suitable for complex problems and long documents.

Ideal Use Cases

Wraith-8B is particularly well-suited for:

Mathematical problem-solving across various domains (arithmetic, algebra, calculus).
STEM tutoring and educational applications requiring detailed, step-by-step reasoning.
Scientific analysis and logical deduction tasks.
Technical writing that benefits from a precise, analytical approach.
Truthful Q&A systems where factual accuracy is paramount.