Modelpublic

sayeed105236/CuttyMOA-1.0

sourceHugging Faceapache-2.0updated 17h agoView on Hugging Face

1likes61downloads

Model Card

🚀 CuttyMOA-1.0 Pro (24K)

Self-trained language model built on Qwen2.5-14B-Instruct, combining two independent LoRA training phases (4K + 20K) via TIES delta merge into a single 24K-step model.

What's New — 20K Merge Update

This release merges both training adapters into one coherent model:

Phase	Platform	Steps	Method	Status
Phase 1	HuggingFace Spaces (L40S 48GB)	~4,000	QLoRA 4-bit nf4, r=128, α=256	✅ Completed
Phase 2	RunPod A40 (48GB VRAM)	~20,000	QLoRA 4-bit nf4, r=128, α=256	✅ Completed
TIES Merge	VPS (12 cores, 47GB RAM)	—	Sign-resolve + magnitude delta combine	✅ Completed
GGUF Q4KM	VPS (llama.cpp)	—	4.87 BPW	✅ Completed
Deploy + HF Upload	VPS + HuggingFace	—	llama-server :5401 + this repo	✅ Live

Why TIES Merge?

—The 20K RunPod phase was a fresh training (not a continuation of 4K)
—Both adapters trained independent learning trajectories on the same base
—TIES (Trim, Elect Sign, Merge) resolves cases where the two adapters modified weights in opposite directions, keeping the larger-magnitude delta — preventing destructive interference
—Result: a combined model that captures knowledge from both phases

Training Data

—204 verified datasets (24 categories) → 4.5M training pairs
—Categories: core LLM, code, security/cyber, Q&A, multilingual (10 lang), reasoning/math, agentic-tools, finance, medical, legal, RAG, enterprise, image, video, audio, presentation, science, sentiment, dialogue, docs/PDF, news, SQL, dialogue systems, translation
—RAG corpus: 857K domain pairs captured from live CUTEADMOA platform usage (MOA fleet, OpenAI-compatible API, security scanning)
—Historical sessions: 2,323 logged interactions

Files

File	Purpose
`cuttymoa-1.0-pro-Q4_K_M.gguf`	Pro model (24K TIES-merged) — recommended
`cuttymoa-1.0-Q4_K_M.gguf`	Original 4K Phase-1 model (legacy)

Deployment

—Self-hosted: llama-server on port 5401 (Q4KM, 8.99 GB, 32K context)
—API: via CUTEADMOA-5.6 MOA fleet (cuttymoa_pro model id)
—Inference: ~30-50 tokens/sec on CPU (12-core VPS)

Capabilities

—General chat & instruction following
—Code generation (Python, JS, SQL, shell)
—Cybersecurity (17 security intents): vulnerability analysis, code audit, threat intel
—Multilingual (10 languages)
—Agentic / tool use
—RAG-augmented domain knowledge

Limitations

—14B parameters, domain-specialized (not frontier-scale)
—Trained on ~50M tokens (far less than commercial models' trillions)
—Best used within the CUTEADMOA Mixture-of-Agents fleet as a specialized engine

📄 Publication

This model is part of the CUTEADMOA platform presented in:

Md Abu Sayeed. CUTEADMOA: A Unified Mixture-of-Agents Platform with Continuous Self-Training for Autonomous Enterprise AI Automation. Advances in Artificial Intelligence and Robotics Research, Vol. 2, No. 3, pp. 190–208, September 2026.

—📰 Paper page: <https://www.scirp.org/journal/paperinformation?paperid=154162>
—📄 Full-text PDF: <https://www.scirp.net/pdf/airr_2970152.pdf>
—🔗 DOI: <https://doi.org/10.4236/airr.2026.23011>

bibtex

@article{sayeed2026cuteadmoa,
  author  = {Md Abu Sayeed},
  title   = {CUTEADMOA: A Unified Mixture-of-Agents Platform with Continuous Self-Training for Autonomous Enterprise AI Automation},
  journal = {Advances in Artificial Intelligence and Robotics Research},
  volume  = {2},
  number  = {3},
  pages   = {190--208},
  year    = {2026},
  issn    = {3143-3987},
  doi     = {10.4236/airr.2026.23011},
  url     = {https://www.scirp.org/journal/paperinformation?paperid=154162}
}

Citation

bibtex

@misc{cuttymoa2026,
  title={CuttyMOA-1.0: Self-Trained 14B Model with TIES Multi-Adapter Merge},
  author={Sayeed},
  year={2026},
  url={https://huggingface.co/sayeed105236/CuttyMOA-1.0}
}