xorfice 0.1.12

SOTA Multimodal Inference Engine (S2S, I2I, V2V) for Xoron-Dev.

🚀 Xoron-Dev: Unified Multimodal AI Model

A state-of-the-art multimodal MoE model that unifies text, image, video, and audio understanding and generation.

Architecture | Features | Installation | Usage | Training | Documentation

---

🏗️ Architecture Overview

Xoron-Dev is built on a modular, mixture-of-experts architecture designed for maximum flexibility and performance.

🧠 LLM Backbone (Mixture of Experts)

• 12 Layers, 1024d, 16 Heads - Optimized for efficient inference and training.
• Aux-Lossless MoE - 8 experts with top-2 routing and configurable shared expert isolation.
• Ring Attention - Memory-efficient processing for up to 128K context.
• Qwen2.5 Tokenizer - High-density 151K vocabulary for multilingual and code support.

👁️ Vision & Video

• SigLIP-2 Encoder - 384px native resolution with multi-scale support (128-512px).
• TiTok 1D Tokenization - Compressed visual representation (256 tokens) for faster processing.
• VidTok 3D VAE - Efficient spatiotemporal video encoding with 4x8x8 compression.
• 3D-RoPE & Temporal MoE - Sophisticated motion pattern recognition and spatial awareness.

🎤 Audio System

• Raw Waveform Processing - Direct 16kHz audio input/output (no Mel spectrograms required).
• Conformer + RMLA - Advanced speech-to-text with KV compression.
• BigVGAN Waveform Decoder - High-fidelity direct waveform generation with Snake activation.
• Zero-Shot Voice Cloning - Clone voices from short reference clips using speaker embeddings.

---

🌟 Features

Multimodal Capabilities

Modality	Input	Output	Strategy
Text	128K Context	Reasoning, Code, Agentic	MoE LLM
Image	128-512px	Understanding & SFT	SigLIP + TiTok
Video	8-24 Frames	Understanding	VidTok + 3D-RoPE
Audio	16kHz Waveform	ASR & TTS	Conformer + BigVGAN

Agentic & Tool Calling

• 250+ Special Tokens for structured agent behaviors.
• Native Tool Use: Execute shell commands, Python scripts, and Jupyter notebooks.
• Reasoning: Advanced Chain-of-Thought (<|think|>, <|plan|>) for complex tasks.
• Safety: Anti-hallucination tokens (<|uncertain|>, <|cite|>) and confidence scores.

Optimization

• LoRA Variants: LoRA+, rsLoRA, and DoRA (r=32, α=64).
• Lookahead Optimizer: Enhanced stability and faster convergence.
• 8-bit Optimization: Save up to 75% optimizer memory with bitsandbytes.
• Continuous-Scale Training: Adaptive resolution sampling for optimal VRAM usage.

---

🚀 Installation

# Clone the repository
git clone https://github.com/nigfuapp-web/Xoron-Dev.git
cd Xoron-Dev

# Install dependencies
pip install -r requirements.txt

---

💻 Usage

Quick Start (Inference)

from load import load_xoron_model

# Load model and tokenizer
model, tokenizer, device, config = load_xoron_model("Backup-bdg/Xoron-Dev-MultiMoe")

# Generate response
output = model.generate_text("Explain quantum entanglement.", tokenizer)
print(output)

CLI Training

The build.py script provides a powerful interface for training and building models.

# Build a new model from scratch
python build.py --build

# Targeted Fine-tuning
python build.py --hf --text --math        # Fine-tune on Math
python build.py --hf --text --agent       # Fine-tune on Agentic tasks
python build.py --hf --video              # Fine-tune on Video datasets
python build.py --hf --voice              # Fine-tune on Audio/Voice

Granular Text Training Flags

Flag	Description
--math	Focus on mathematical reasoning and steps.
--agent	Tool use, code execution, and system operations.
--software	High-quality software engineering and coding.
--cot	Chain-of-Thought and logical reasoning.
--medical	Medical knowledge and clinical reasoning.
--hallucination	Anti-hallucination and truthfulness.

---

🏋️ Training

Weighted Loss Strategy

The trainer applies specialized weights to ensure high performance on critical tokens:

• Reasoning (CoT): 1.5x
• Tool Calling: 1.3x
• Anti-Hallucination: 1.2x

Continuous-Scale Strategy

Xoron-Dev dynamically samples resolutions during training:

• Image: 128px to 384px (step=32)
• Video: 8 to 24 frames, 128px to 320px

---

📦 Export & Quantization

Export your models for efficient deployment:

# Export to GGUF (for llama.cpp)
python build.py --hf --gguf --gguf-quant q4_k_m

# Export to ONNX
python build.py --hf --onnx --quant-bits 4

---

🤝 Contributing

Contributions are welcome! If you have ideas for new modalities or optimizations, please open an issue or PR.

---

📄 License

This project is licensed under the MIT License.

---

Built with ❤️ by the Xoron-Dev Team