YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SO8T-AEGIS-phi3.5-v3.0 (v3.0.0)
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.
| Metric | Value | Significance |
|---|---|---|
| ANOVA (F-value) | N/A | N/A |
| p-value | N/A | Not significant |
| Cohen's d | N/A | Large Effect Size (>0.8) |
| 95% CI | [N/A, N/A] | - |
Japanese-Specific Benchmarks
| Benchmark | AEGIS-v3.0 | Base (Borea) | Diff |
|---|---|---|---|
| ELYZA-100 | TBD | 4.2 | +N/A |
| J-MMLU | TBD | 0.65 | +N/A |
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:
- Sakana AI (2025): Auto-Retraining Pipeline for Intelligent Agents.
- DeepSeek (2025): GRPO: Group Relative Policy Optimization for LLMs.
- Borea (2024): Japanese-English Bilingual LLM Optimization.
- 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.