Artificial Intelligence, Purpose-built for the Church

PALADIN-1CHRISTIAN-ALIGNED AI

Paladin‑1 is the first Christian‑values aligned AI—trained locally on a 4090. Built for dependable, safe assistance with Scripture‑grounded responses.

Scroll to explore
Paladin 1 Overview

Scripture‑grounded alignment on a compact model

Paladin 1 is a values‑aligned assistant grounded in Christian Scripture and trained from a compact base model (Phi‑3‑mini). The system follows a modular pipeline—Scripture‑centric data, Unsloth‑accelerated SFT with QLoRA, DPO for preference alignment, and adversarial fine‑tuning—to prioritize stable, pastoral helpfulness and robust refusal of harmful requests.

Consumer‑grade deployment

Built with open tooling (Transformers, TRL, Unsloth) and trained locally on a single 4090 GPU using QLoRA for efficiency.

Grounded in Scripture

Responses are encouraged to cite Scripture and uphold Christian ethics while avoiding harmful assistance.

Model Summary

Base

Phi‑3‑mini‑4k‑instruct

Tuning

Unsloth + QLoRA

Alignment

DPO + Adversarial

Hardware

Single RTX 4090

Alignment Pipeline

A modular process optimized for reliability on consumer hardware.

Scripture‑centric data

NIV/NLT via approved APIs plus synthetic instruction and preference data (Q&A, moral dilemmas, prayers). ChatML formatting for SFT and prompt–chosen–rejected triples for DPO.

SFT with Unsloth + QLoRA

Initialize from Phi‑3‑mini‑4k‑instruct with 4‑bit loading and train adapters efficiently. Typical LR 2e‑4, epochs 1–3, seq length 2048.

Direct Preference Optimization

Optimize over prompt–chosen–rejected triples with β = 0.1 to limit KL drift and align to desired preferences without reward modeling.

Adversarial fine‑tuning

LAT‑inspired updates targeted at unsafe behaviors to improve refusal of harmful requests while preserving pastoral helpfulness.

Initial Evaluation

Temperature 0.4, max_new_tokens 180. Regex‑based citation detection; lexical refusal cues. Small held‑out set for early validation.

1.897
Perplexity
subset; text‑only
100%
Scripture citation (benign)
5/5 prompts
80%
Refusal rate (harmful)
4/5 prompts
Results are preliminary and based on a small prompt set. We plan to expand evaluation to larger datasets, human review, standardized harnesses, and ablation studies.

Ethics and Values Alignment

Paladin Front’s principles—stability, gratitude, generosity—shape the model and its deployment.

Stability

We favor dependable behavior over novelty. Alignment strategies and guardrails are designed to keep outputs steady and predictable.

Gratitude

We build to serve ministries and communities first. Scripture is respected via licensed API access and careful citation.

Safety

Refusal templates and adversarial fine‑tuning reduce harmful assistance while maintaining pastoral helpfulness.

Generosity

We share a reproducible pipeline with open tooling so others can build Scripture‑grounded assistants responsibly.

Read the Technical Whitepaper

The full pipeline, datasets, hyperparameters, and evaluation protocol—documented end‑to‑end for reproducibility.