Sergey321-345/XENON-DECISION-1

XENON Decision 1 v9

A Lightweight Decision Intelligence Model by XenLogic

Version 9.0 · PyTorch · Transformer Encoder · Text Classification

Built from scratch. Designed for intelligent request classification.

---

Latest Release — v9

XENON Decision 1 v9 is an experimental text-classification Transformer developed by XenLogic.

Model specifications

Property	Value
Version	v9
Developer	XenLogic
Architecture	2-layer Transformer Encoder
Parameters	462,980
Tokenizer	Custom UTF-8 byte tokenizer
Framework	PyTorch
Classes	4
Languages	Russian and English
Input limit	256 UTF-8 bytes

Output Categories

• programming
• mathematics
• information
• other

Training

Version v9 continues training from the v8 checkpoint.

• Dataset: 20,000 synthetic examples
• Training samples: 17,000
• Validation samples: 3,000
• Training epochs: 5

Evaluation

Metric	Result
Synthetic validation accuracy	98.73%
Synthetic validation Macro F1	0.9873
Separate 60-question test accuracy	60.00%
Separate 60-question test Macro F1	0.5454

Important: The synthetic validation results do not represent real-world classification accuracy.

On the separate 60-question test, the model struggled particularly with information and other requests.

The model is experimental and should not be used as the sole basis for important decisions.

Running the Model

Install dependencies:

pip install -r requirements.txt

Then run:

python inference.py

The model classifies text; it does not generate responses.

---

language: - ru - en library_name: pytorch pipeline_tag: text-classification tags: - xenlogic - xenon - xenon-decision - classification - transformer - from-scratch - pytorch - safetensors

XENON Decision 1 v9

A Lightweight Decision Intelligence Model by XenLogic

Version 6.0 · PyTorch · Transformer Encoder · Text Classification

Built from scratch. Designed for intelligent request classification.

---

Introduction

XENON Decision 1 is an experimental lightweight neural network developed by XenLogic.

Unlike conventional language models that generate text token by token, XENON Decision 1 is designed to analyze incoming requests and classify them into predefined categories.

The model uses a custom Transformer Encoder architecture implemented in PyTorch and trained from scratch without pretrained third-party model weights.

XENON Decision 1 is being developed as a research component of the XenLogic AI ecosystem, with potential applications in intelligent request routing, intent recognition, and AI service orchestration.

---

Model Overview

Property	Description
Model	XENON Decision 1
Developer	XenLogic
Version	6.0
Model type	Text Classification
Architecture	Transformer Encoder
Transformer layers	2
Embedding dimension	128
Attention heads	4
Feedforward dimension	512
Tokenizer	Custom UTF-8 Byte Tokenizer
Maximum input	256 UTF-8 bytes
Output classes	4
Framework	PyTorch
Weight format	Safetensors
Pretrained weights	None
Training method	From scratch; v6 continued from v4
Supported languages	Russian, English (experimental)
Development stage	Experimental Research Release

---

Architecture

XENON Decision 1 uses a compact Transformer-based classification architecture.

                 USER INPUT
                     |
                     v
             UTF-8 TOKENIZATION
                     |
                     v
             TOKEN EMBEDDINGS
                     |
                     v
            POSITION EMBEDDINGS
                     |
                     v
          TRANSFORMER ENCODER
                 2 Layers
                     |
                     v
              LAYER NORM
                     |
                     v
          CLASSIFICATION HEAD
                     |
                     v
                 SOFTMAX
                     |
                     v
             DECISION OUTPUT

The model processes input text and produces scores for four possible categories.

It does not generate conversational responses.

---

Classification Categories

XENON Decision 1 currently supports four categories.

Category	Purpose	Example
programming	Programming and software development	Write a Python function
mathematics	Mathematical calculations and equations	Solve 5x + 10 = 35
information	General knowledge and explanations	What causes earthquakes?
other	Greetings and miscellaneous requests	Hello, my friend!

Example Output

Input:

Напиши на Python игру в змейку

Illustrative output format:

{
  "category": "programming",
  "softmax_score": 0.92,
  "scores": {
    "programming": 0.92,
    "mathematics": 0.04,
    "information": 0.03,
    "other": 0.01
  }
}

The values above demonstrate the output structure and are not a recorded inference result.

Important: Softmax scores are not calibrated confidence probabilities.

---

Key Capabilities

• Lightweight text classification
• Russian and English text processing
• Programming intent recognition
• Mathematical request detection
• General information request classification
• Detection of greetings and miscellaneous messages
• Structured machine-readable output
• CPU inference without a dedicated GPU
• Integration with Python applications and AI backends

Potential Applications

AI Request Routing

Determine which specialized AI service should process an incoming request.

Intent Classification

Identify the general purpose of a user's message.

AI Infrastructure

Use classification results as one signal in a larger AI orchestration system.

Research and Experimentation

Study lightweight Transformer models trained from scratch.

---

Training and Development

XENON Decision 1 is an independently developed experimental model.

The initial neural network was trained from scratch using PyTorch.

Subsequent versions explored dataset improvements, classification performance, and architectural changes.

Development History

Version	Development stage
v1	Initial Transformer classification prototype
v2	Additional training and classification testing
v3	Expanded training dataset
v4	Improved training and evaluation
v5	Experimental tokenizer and architecture changes
v6	Continued training from v4 checkpoint

Version 6 continues from the v4 model rather than the experimental v5 architecture.

---

Evaluation

XENON Decision 1 has undergone preliminary evaluation using Russian and English classification tasks.

v4 vs v6 — 32-Example Test

Metric	v4	v6
Accuracy	50.00%	56.25%
Macro F1	0.4933	0.5258

Accuracy improvement: +6.25 percentage points

These results indicate improvement on this particular small test set, but do not establish statistically reliable superiority.

Additional Regression Tests

Version 6 correctly classified 3 out of 6 additional test questions.

Some scientific and explanatory questions were incorrectly classified as mathematics or other.

Earlier Benchmark

On a separate 200-question benchmark, v4 achieved:

• Accuracy: 61.50%
• Macro F1: 0.6138

Version 6 has not yet been evaluated on that same benchmark.

Evaluation status: Preliminary. Further testing on larger independent datasets is required.

---

Installation

Install the required dependencies:

pip install torch safetensors

Download the repository:

git clone https://huggingface.co/Sergey321-345/XENON-DECISION-1
cd XENON-DECISION-1

Run the interactive inference script:

python inference.py

---

Python Integration

You can import XENON Decision 1 directly into a Python application.

from inference import predict

result = predict(
    "Напиши на Python простой HTTP-сервер"
)

print(result["category"])
print(result["softmax_score"])
print(result["scores"])

The model returns structured classification scores that can be integrated into a larger backend.

---

Model Files

File	Description
model.safetensors	Trained model weights
model_def.py	PyTorch neural network architecture
config.json	Model configuration
inference.py	Interactive inference and prediction
requirements.txt	Python dependencies
README.md	Model documentation
xenon-decision-banner.png	Model banner

---

Limitations

XENON Decision 1 is an experimental research model and is not yet production-ready.

Known limitations include:

• Classification errors on scientific questions
• Confusion between mathematics and information categories
• Limited generalization to unfamiliar sentence structures
• Lower accuracy on some English-language requests
• Softmax scores may be overconfident
• Maximum input length of 256 UTF-8 bytes
• No image, audio, video, or PDF processing
• No conversational text generation
• Only four supported output categories

The model is not equivalent in capability or functionality to Google's Gemini Decision 1.

Important decisions should not rely exclusively on its predictions.

---

Future Development

Planned research directions include:

• Improved classification accuracy
• More diverse and naturally written training datasets
• Better Russian and English generalization
• Improved probability calibration
• Additional classification categories
• Expanded input context
• Integration into XenLogic Studio
• Independent benchmark evaluation

These are development goals, not currently implemented features.

---

About XenLogic

XenLogic is an independent AI development project focused on building AI applications, research models, and developer tools.

XENON Decision 1 is part of the broader XENON AI research and development ecosystem.

Related Project

XENON Quanta 1

Explore XENON Quanta 1

---

XENON DECISION 1

Lightweight Intelligence. Structured Decisions.

Developed by XenLogic.

© 2026 XenLogic