Paterikon-SFT-v1

Name: paterikon-sft-v1
Availability: InStock

Single-pass supervised fine-tune of Paterikon-3B on Orthodox Q&A pairs.

Overview

Paterikon-SFT-v1 is the first instruction-tuned checkpoint derived from Paterikon-3B. It was fine-tuned with supervised fine-tuning (SFT) on ~140M tokens of synthetically-generated Orthodox Christian Q&A pairs covering theology, spiritual life, Church Fathers, liturgy, and pastoral guidance.

The base model (Paterikon-3B) already had patristic domain fluency from continued pre-training on 116M tokens of Church Father writings. This SFT pass teaches it to answer questions in a pastoral Orthodox voice with theological accuracy.


Base model	jayfurzy/paterikon-3b (Qwen2.5-3B CPT)
Training	Full fine-tune SFT (TRL)
Parameters	3.09 billion
Architecture	Qwen2ForCausalLM, 36 layers, 2048 hidden
Languages	English (primary), Russian, Greek
Training data	~140M tokens synthetic Orthodox Q&A pairs
Frameworks	TRL 0.28.0, Transformers 4.57.6, PyTorch 2.9.1
License	Apache 2.0

Quick Start

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "jayfurzy/paterikon-sft-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are an Orthodox Christian theologian, answering in the spirit of the Holy Fathers."},
    {"role": "user", "content": "What does St. John Chrysostom teach about wealth and poverty?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Intended Use

Orthodox theological Q&A
Pastoral guidance and spiritual counsel
Patristic text generation and completion
Educational tool for Orthodox Christian studies
Base for further fine-tuning (RLHF, DPO)

Limitations

May confabulate on specific factual questions about Church history dates, canon law, or exact patristic citations
Not a substitute for a spiritual father or priest
Reflects the synthetic data distribution — heavy on Russian Orthodox sources
Temperature > 0.8 may produce theologically inconsistent outputs

Training Pipeline

Qwen2.5-3B-Instruct → CPT (Paterikon-3B) → SFT (this model) → [iterative SFT, DPO planned]

Citation

bibtex

@misc{paterikon-sft-v1,
  author = {Justin Fursov},
  title = {Paterikon-SFT-v1: Orthodox Patristic Instruction-Tuned Language Model},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/jayfurzy/paterikon-sft-v1},
}