logos-v61-sft

Logos v6.1-SFT — Abrahamic Religious Texts Model (Research Alpha)

RESEARCH USE ONLY — This model has known limitations. Not suitable for religious advice, fatwa generation, or authoritative scriptural interpretation.

Overview

Logos v6.1-SFT is a research-grade language model specialized in Abrahamic religious texts across Arabic, Hebrew, Greek, English, Latin, fine-tuned from Qwen2.5-7B-Instruct using QLoRA (r=32, alpha=64).

This version includes:

• Base LM training: 3 epochs on 130K verses (multilingual corpus)
• SFT training: 30K instruction-tuned Q&A samples
• Citation SFT: Additional 1 epoch on 8.2K citation-format examples (experimental)

Model Details

Training Data

Base Corpus (130K verses)

SFT Dataset (30K samples)

Instruction-tuned Q&A pairs covering verse recall, cross-tradition comparison, citation lookup, tafsir explanation, and multilingual text understanding.

Citation SFT Dataset (8.2K samples)

Citation formatting examples for Quran (Surah X:Y), Bible (Book Chapter:Verse), and Talmud (tractate page) formats.

Evaluation (v6.1-SFT, seed=42, greedy)

Known Limitations

• Low citation accuracy (20%) — inconsistent surah:ayah and book:chapter:verse formatting
• Zero exact verse recall — model has not memorized verse text; relies on RAG for verbatim quotes
• Hallucination rate 50% — too high for production factual queries
• Tafsir not learned — needs DPO or targeted tafsir training
• Arabic diacritics — decent F1 but not scholarly-grade tashkeel
• Research only — not intended for religious advice or authoritative interpretation

Usage

Transformers (GPU)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "frostedunicorn/logos-v61-sft",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("frostedunicorn/logos-v61-sft")

prompt = "Explanation of Surah 1:1 in the Quran:"
messages = [{"role": "user", "content": prompt}]
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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Ollama

ollama create logos-v61-sft -f Modelfile
ollama run logos-v61-sft "What is the first surah of the Quran?"

Chat Template

{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
{{ end }}{{ .Response }}<|im_end|>

System Prompt

You are Logos, an AI assistant specialized in Abrahamic religious texts across Hebrew, Greek, Arabic, English, and Latin. You provide accurate citations from Torah, Bible, Quran, and related sources. You handle sacred texts with reverence and never mix registers inappropriately. When you don't know something, you say so.

License & Attribution

Model weights: Apache 2.0 (research use)

Upstream data licenses:

• Quran: Public Domain
• WLC Hebrew Bible: Public Domain
• Septuagint (LXX): Public Domain
• Talmud/Mishnah (Rodkinson): Public Domain
• Hadith English translations: Various (CC-BY, CC-BY-SA)
• Sefaria texts: CC-BY / CC-BY-SA (requires attribution)
• Gutenberg texts: Public Domain

Citation

@misc{logos2026,
  title={Logos v6.1-SFT: Abrahamic Religious Texts Model},
  author={dtfrost5},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/frostedunicorn/logos-v61-sft}}
}

---

Evaluated 2026-09-17. Training loss: 0.421, combined SFT + citation LoRA merged.