AEGIS-phi3.5-jp-v3.0

SO8T-AEGIS-phi3.5-v3.0 (v3.0.0)

[!IMPORTANT] This model is an advanced iteration of the AEGIS series, retrained on Borea-Phi-3.5 with SO8T Quadrality and Sakana AI hybrid research integration.

Overview

AEGIS-v3.0 is a specialized Large Language Model (LLM) designed for Scientific Discovery, OSINT Intelligence, and National Security Analysis. It leverages a unique quadrality reasoning framework (SO8T) to ensure logical rigor, safety, and policy compliance.

Statistical Benchmark Analysis (Phase 6)

We conducted a comprehensive evaluation using a multi-phase statistical framework.

Japanese-Specific Benchmarks

Key Technologies

• SO8T Quadrality Reasoning: Structured thinking with <think-task>, <think-analysis>, <think-safety>, and <think-policy>.
• DeepSeek-style GRPO: Group Relative Policy Optimization for enhanced mathematical and OSINT extraction reasoning.
• mHC Manifold Integration: Manifold Harmonic Correction to stabilize weights during LoRA/QLoRA adaptation.
• Unsloth & imatrix: Ultra-fast training and high-precision quantization for local inference (RTX 3060 optimized).

Scientific Citations & Assets

This model incorporates methodologies and data from:

1. Sakana AI (2025): Auto-Retraining Pipeline for Intelligent Agents.
2. DeepSeek (2025): GRPO: Group Relative Policy Optimization for LLMs.
3. Borea (2024): Japanese-English Bilingual LLM Optimization.
4. SO8T Paper (2025): Quadrality Reasoning in Adaptive AI Systems.

Training Environment

• Hardware: NVIDIA RTX 3060 (12GB VRAM)
• Software: Unsloth, Transformers, TRL, bitsandbytes
• Dataset: Integrated 2024-2026 World Events, National Security Documents, and Scientific Arxiv/BioRxiv papers.

Disclaimer

This model is for research and scientific discovery. Users should verify OSINT information with external sources.