Prycat/Prycat1

Prycat 1

First public release of Prycat -- a QLoRA fine-tune of Hermes-3-Llama-3.1-8B trained via the BlackLoRA-N closed-loop self-improvement system.

What is Prycat?

Prycat is trained through a closed feedback loop:

1. The model generates responses
2. A blended LLM + rule-based judge scores them across accuracy, logic, tone, and safety
3. High-scoring pairs are selected as training data
4. QLoRA fine-tuning anchors the weights to those pairs
5. A probe suite (axiom interview) lets the human set immutable ground truth the model cannot game
6. A drift monitor detects if the judge and student start co-evolving and aborts training

This release

• Base: NousResearch/Hermes-3-Llama-3.1-8B
• Method: QLoRA (4-bit, rank 16, alpha 32)
• Training steps: 200
• Training loss: 0.256
• System: BlackLoRA-N v1

Files

• dapters/latest/ -- raw LoRA adapter (load with PEFT)
• dapters/merged/ -- merged 16-bit weights (load directly with transformers)

Versioning

Prycat1 = this drop. Each subsequent BlackLoRA-N training cycle produces a new numbered release (Prycat2, Prycat3...).