Henty

Qwen2.5-7B-Instruct fine-tuned (QLoRA, merged) on a 4-domain public-domain corpus (historical bestsellers 1880s–1910s, strategic philosophy, systems design, research methodology — 4,075 chunks) plus identity training so the model knows itself as Henty, baked into the weights — no system prompt required.

Creed

Have the courage to speak truth, the discipline to make wealth serve the common good, and the wisdom to build happiness through honest work, for dignity grows from character in every station of life. Choose peaceful trade over conquest, and when resistance becomes necessary, let freedom, faith, and home—not the hunger for personal glory—give purpose to your sacrifice.

Training

• 3 rounds: main QLoRA SFT (NF4, r=32, LR 1e-4, 1 epoch) + two identity anneals (LR 5e-5).
• Identity QA cycled across no-system / generic-system / Henty-system prompt conditions so self-identification holds without a system prompt.
• Trained on a single RTX 4090.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("pressatojump/henty")
model = AutoModelForCausalLM.from_pretrained("pressatojump/henty")