tritter-500m-hybrid-bitnet

Tritter 500M Hybrid BitNet

Model Details

• Model Size: 500M parameters
• Architecture: BitNet 1.58-bit ternary quantization
• Training Methodology: Hybrid Predictive Training (Embedding-Prediction Paradigm)
• Quantization: {-1, 0, 1} ternary weights

Overview

This model is trained using Hybrid Predictive Training, which combines:

• Embedding-prediction paradigm: Core computation in continuous embedding space
• BitNet 1.58-bit quantization: Efficient ternary weight representation
• Dual prediction heads: Both embedding and token space outputs during training

The model operates in continuous embedding space at inference time, with token prediction as temporary scaffolding for training compatibility.

Architecture Specifications

• Hidden Size: 2048
• Number of Layers: 16
• Number of Attention Heads: 32
• Intermediate Size: ~3.5x hidden (7168)
• Max Position Embeddings: 4096
• Context Window: 4K tokens

Training Data

• Total Tokens: ~100B tokens
• Data Mix: Code-centric (Python, Rust, technical documentation)
• Quality Gates: Hardcoded secrets rejected, security checks enabled

Comparison with Standard Training

For comparison with the standard trained baseline, see:

• Standard 500M: tzervas/tritter-500m-bitnet

Key Differences:

Training Metrics

Convergence Comparison:

Final Metrics:

• Final Training Loss: pending
• Final Validation Loss: pending
• Training Time: pending
• Hardware: RTX 5080 16GB (with gradient checkpointing)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("tzervas/tritter-500m-hybrid-bitnet")
tokenizer = AutoTokenizer.from_pretrained("tzervas/tritter-500m-hybrid-bitnet")

# Generate text
inputs = tokenizer("def fibonacci", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Model Details

• Framework: PyTorch
• Quantization Method: BitNet 1.58-bit ternary
• License: MIT

Research Background

For more information on hybrid predictive training, see:

• Embedding-Prediction Paradigm: Operating in continuous embedding space
• BitNet 1.58-bit: Efficient ternary quantization {-1, 0, 1}
• Progressive Layer Loading: Support for larger models on limited VRAM

Citation

If you use this model, please cite:

@model{tritter500m_hybrid,
  author={Tzervas, K.},
  title={Tritter 500M Hybrid BitNet: Embedding-Prediction Training},
  year={2025},
  publisher={Hugging Face}
}

---

Created as part of the Tritter multimodal transformer research project.