CuttyMOA-1.0

CuttyMOA-1.0 is the first self-trained model from the CUTEADMOA platform. It is a Qwen2.5-14B-Instruct base fine-tuned with a 4K LoRA adapter (QLoRA 4-bit, nf4, r=128) for ~4,000 steps on the 204-dataset corpus covering development, security/cyber, code, Q&A, multilingual, reasoning, finance, medical, legal, agentic-tools, RAG, and more.

This repo ships the GGUF Q4_K_M quantization for local/llama.cpp deployment.

Model Details

• Base model: Qwen/Qwen2.5-14B-Instruct
• Fine-tune method: QLoRA 4-bit (nf4, r=128, α=256), SFTTrainer
• Training: ~4,000 steps, HF Spaces L40S (48GB)
• Context length: 32,768 tokens
• Quantization: GGUF Q4_K_M (4.87 BPW), 8.99 GB
• Architecture: qwen2, 48 layers, 40 heads, 8 KV heads, ff 13824

Usage (llama.cpp / llama-server)

# Run local server
llama-server -m cuttymoa-1.0-Q4_K_M.gguf --port 5401 -c 8192

# Chat via OpenAI-compatible API
curl http://localhost:5401/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Hello!"}]}'

Training Data

Trained on the 204 verified datasets (sayeed105236/cuteadmoa-204-datasets) — a corpus of 4.5M training pairs across 24 categories including development, image/video/presentation/audio metadata, security, multilingual and enterprise/RAG content.

Performance

Verified live on deployment: generates coherent domain-aware responses for cybersecurity and code tasks with base knowledge intact (32K ctx).

Roadmap

• CuttyMOA-1.0 Pro: TIES-merge of the 4K adapter with the 20K adapter (RunPod) for enhanced capability.