Technical Report 2025

Task-Specialized Micro Language Models Outperform Larger Zero-Shot Models on Structured Data Extraction

maaza model family

Small language models designed for structured JSON generation on edge devices. Task-specific models that outperform larger general-purpose models on structured output.

Key Results

  • 72.3% JSONExact accuracy, 0.878 Field F1
  • 70% tool routing accuracy, 70ms latency
  • CPU-only inference, <150MB disk footprint

Download Paper

Version 0.7

Dataset Zenodo 2025

A Large-Scale ML-Guided Search for 28-Term Prime Progressions

Computational search for arithmetic progressions of primes with 28 or more terms. ML-guided candidate generation across 10^9 candidates found no progressions exceeding 10 primes.

Security Research

Protecting AI infrastructure through independent security research.

  • Hugging Face
  • Microsoft ONNX Runtime
  • ONNX
  • Ollama
  • llama.cpp
  • OpenRouter
  • LiteLLM
  • LlamaIndex
  • Triton Inference Server
  • H2O
  • OpenMMLab
  • NLTK
  • Fairseq
  • Keras Tuner

Citations

Task-Specialized Micro Language Models...
@techreport{maaza2025,
  title={Task-Specialized Micro Language Models Outperform Larger
         Zero-Shot Models on Structured Data Extraction},
  author={CycleCore Technologies},
  institution={CycleCore Technologies},
  year={2025},
  type={Technical Report},
  url={https://cyclecore.ai/research}
}
ML-Guided Search for 28-Term Prime Progressions...
@dataset{cyclecore_ap28_2025,
  author={CycleCore Technologies},
  title={A Large-Scale ML-Guided Search for 28-Term Prime
         Progressions: No Progressions with More Than 10
         Primes Found Among 10^9 Candidates},
  year={2025},
  publisher={Zenodo},
  version={v1.0},
  doi={10.5281/zenodo.17889361},
  url={https://doi.org/10.5281/zenodo.17889361}
}