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?
--- language: - en license: llama3 base_model: NousResearch/Hermes-3-Llama-3.1-8B tags: - lora - qlora - fine-tuned - prycat - blackwell --- # 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...).
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?
