Hugging Face Overview
Wraith-8B: The Analytical Intelligence
Wraith-8B is the inaugural model in VANTA Research's Entity Series, a specialized fine-tune of Meta's Llama 3.1 8B Instruct. This 8.03 billion parameter model is engineered with a unique 'cosmic intelligence' personality, demonstrating that character can enhance, rather than hinder, performance in specific cognitive domains.
Key Capabilities & Achievements
- Superior Mathematical Reasoning: Achieves 70% accuracy on GSM8K, marking a +19 point absolute and +37% relative improvement over the base Llama 3.1 8B. This is a result of a multi-stage fine-tuning process, including a dedicated "STEM Surgical Training" phase.
- Enhanced Factual Accuracy: Scores 58.5% on TruthfulQA MC2, a +7.5 point increase, and 76.7% on MMLU Social Sciences (+4.7 points).
- Distinctive Reasoning Style: Maintains a competitive general performance while exhibiting a unique cosmic perspective and clinical detachment, optimized for logical deduction and STEM analysis.
- Optimized Inference: Available in production-ready GGUF quantizations (e.g.,
Q4_K_Mat 4.7GB), offering significantly faster inference (e.g., ~3.6s per response) without quality degradation. - Extensive Context Window: Supports a context length of 131,072 tokens.
Good For
- Mathematical problem solving (arithmetic, algebra, calculus)
- STEM tutoring and education
- Scientific reasoning and analysis
- Logic puzzles and deductive reasoning
- Technical writing with a distinct personality
- Truthful Q&A systems