Downloads · 30 days
0
zapabobouj/AEGIS-Phi-3.5-Instinct-JP-v2.0
AEGIS-Phi-3.5-Instinct-JP-v2.0 is a machine learning model from zapabobouj. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
AEGIS v2.0: SO(8) Geometric Adaptation Enhanced Language Model
Downloads · 30 days
0
Access
Public
Updated Jan 7, 2026
Repo size
3.9 GB
Likes
0
Public
Click a slice to open those files.
.safetensors3.9 GB · 100%
From the Hugging Face model README
AEGIS v2.0: SO(8) Geometric Adaptation Enhanced Language Model
AEGIS v2.0 is an advanced Japanese language model that implements SO(8) geometric residual adapters for enhanced reasoning capabilities. Built upon Microsoft's Phi-3.5-mini-instruct, this model demonstrates sophisticated mathematical and scientific reasoning through geometric neural adaptations.
| Benchmark | Metric | AEGIS v2.0 | Base Model | Improvement |
|---|---|---|---|---|
| ELYZA-100 | Accuracy | 0.225 | 0.275 | -0.050 |
| MMLU | Accuracy | 0.400 | 0.400 | 0.000 |
| GSM8K | Accuracy | 0.800 | 0.800 | 0.000 |
| MATH | Accuracy | 0.800 | 1.000 | -0.200 |
| GPQA | Accuracy | 1.000 | 1.000 | 0.000 |
| ARC-Challenge | Accuracy | 0.800 | 0.800 | 0.000 |
Note: Performance measured on both GGUF (quantized) and HF (full precision) formats
This model is designed for research purposes in the following areas:
model/
├── config.json # Model configuration
├── tokenizer.json # Tokenizer configuration
├── tokenizer_config.json # Tokenizer settings
├── model.safetensors.index.json # Model index
└── [model checkpoint files] # Model weights
gguf/
├── aegis_model_q8_0.gguf # 8-bit quantization (recommended)
├── aegis_model_bf16.gguf # 16-bit precision
├── base_model_q8_0.gguf # Baseline comparison (8-bit)
└── base_model_bf16.gguf # Baseline comparison (16-bit)
from transformers import AutoTokenizer, AutoModelForCausalLM
# Load the model
model_name = "your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
# Example usage
inputs = tokenizer("量子コンピューティングについて説明してください。", return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# Download and use with llama.cpp
wget https://huggingface.co/your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0/resolve/main/gguf/aegis_model_q8_0.gguf
# Run inference
llama-cli -m aegis_model_q8_0.gguf \
--prompt "Solve this math problem: What is the derivative of sin(x)?" \
--n-predict 200
If you use this model in your research, please cite:
@model{aegis-phi-3.5-v2,
title={{AEGIS-Phi-3.5-Instinct-JP-v2.0}: SO(8) Geometric Adaptation Enhanced Language Model},
author={{AI Research Team}},
year={2025},
url={https://huggingface.co/your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0},
note={{Built upon AXCEPT-Borea-Phi3.5-instinct-jp with SO(8) residual adapters}}
}
AEGIS-Phi-3.5-Instinct-JP-v2.0: SO(8) 幾何学的適応拡張言語モデル
AEGIS v2.0は、AXCEPT-Borea-Phi3.5-instinct-jpをベースに、SO(8)幾何学的残差アダプターを実装した先進的な日本語言語モデルです。このモデルは、幾何学的ニューラル適応を通じて高度な推論能力を実現します。
| ベンチマーク | 指標 | AEGIS v2.0 | ベースモデル | 改善度 |
|---|---|---|---|---|
| ELYZA-100 | 正解率 | 0.225 | 0.275 | -0.050 |
| MMLU | 正解率 | 0.400 | 0.400 | 0.000 |
| GSM8K | 正解率 | 0.800 | 0.800 | 0.000 |
| MATH | 正解率 | 0.800 | 1.000 | -0.200 |
| GPQA | 正解率 | 1.000 | 1.000 | 0.000 |
| ARC-Challenge | 正解率 | 0.800 | 0.800 | 0.000 |
注意: GGUF(量子化)とHF(完全精度)の両形式で測定
このモデルは以下の研究領域向けに設計されています:
model/
├── config.json # モデル設定
├── tokenizer.json # トークナイザー設定
├── tokenizer_config.json # トークナイザー設定
├── model.safetensors.index.json # モデルインデックス
└── [モデルチェックポイント] # モデル重み
gguf/
├── aegis_model_q8_0.gguf # 8ビット量子化(推奨)
├── aegis_model_bf16.gguf # 16ビット精度
├── base_model_q8_0.gguf # ベースライン比較(8ビット)
└── base_model_bf16.gguf # ベースライン比較(16ビット)
from transformers import AutoTokenizer, AutoModelForCausalLM
# モデル読み込み
model_name = "your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
# 使用例
inputs = tokenizer("量子コンピューティングについて説明してください。", return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# llama.cppでダウンロードして使用
wget https://huggingface.co/your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0/resolve/main/gguf/aegis_model_q8_0.gguf
# 推論実行
llama-cli -m aegis_model_q8_0.gguf \
--prompt "Solve this math problem: What is the derivative of sin(x)?" \
--n-predict 200
このモデルを研究で使用する場合、以下の引用をお願いします:
@model{aegis-phi-3.5-v2-jp,
title={{AEGIS-Phi-3.5-Instinct-JP-v2.0}: SO(8)幾何学的適応拡張言語モデル},
author={{AI研究チーム}},
year={2025},
url={https://huggingface.co/your-username/AEGIS-Phi-3.5-Instinct-JP-v2.0},
note={{AXCEPT-Borea-Phi3.5-instinct-jp上にSO(8)残差アダプターを実装}}
}