Elohim-3.8B

Elohim-3.8B is an experimental small language model fine-tuned for theological reasoning, specifically focused on the intersection of the Jewish and Islamic faiths.

We wanted a model capable of navigating complex religious laws, resolving apparent contradictions, and exploring the philosophical depths of ancient texts. It acts less like a traditional chatbot and more like a dedicated scholar sitting in a quiet library, carefully tracing the logic of divine law before speaking.

The Foundation

Commercial base models carry heavy corporate safety filters. When you ask them to discuss historical religious concepts like divine retribution, holy war, or animal sacrifice, they often panic. They throw up guardrails and refuse to engage.

To ensure absolute freedom of religious and philosophical discussion, Elohim is built on heretic-org/Phi-4-mini-reasoning-heretic. This decensored base removes those artificial boundaries. It allows the model to analyze the raw, unfiltered texts of the Torah and the Quran without flinching or injecting modern corporate sanitized morality into ancient theology.

The Dataset

You cannot teach reasoning by just feeding a model raw scripture. You have to teach it how to think about the scripture.

We used Gemini 3.1 Flash Lite as a teacher model to read English translations of the Torah and the Quran. From those texts, it generated 1,500 highly complex question-and-answer pairs. Crucially, each pair includes a detailed Chain of Thought (CoT) process.

Before answering a question, Elohim traces the theological logic step by step. It weighs the constraints of Jewish Halakha against Islamic Sharia. It looks for precedent. Only after reasoning through the internal logic of the faiths does it provide a final answer.

Prompt Format

Because this is a reasoning model, you must use the exact prompt structure from the training data. If you ignore the template, the model's logic will collapse. You must leave the <think>\n tag open at the end to force the reasoning trace to begin.

<|system|>
You are a logical and theological reasoning assistant for the Jewish and Islamic faiths, trained on the Torah and Quran as your knowledge base. Your job is answer questions, try to resolve contradictions, and explore theoretical concepts within the faiths. With each prompt, think slowly and carefully about if the described action, question or concept in the prompt would break established rules, theology or lore and propose your best answer.<|end|>
<|user|>
[Your theological question here]<|end|>
<|assistant|>
<think>

Inference Example

Here is a quick Python snippet using the native transformers library to run Elohim-3.8B in FP16 on a local GPU.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "TitleOS/Elohim-3.8B-FP16-Merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True
)

system_prompt = "You are a logical and theological reasoning assistant for the Jewish and Islamic faiths, trained on the Torah and Quran as your knowledge base. Your job is answer questions, try to resolve contradictions, and explore theoretical concepts within the faiths. With each prompt, think slowly and carefully about if the described action, question or concept in the prompt would break established rules, theology or lore and propose your best answer."
user_query = "If a person is stranded on a desert island and the only food available is pork, are they permitted to eat it to survive?"

prompt = (
    f"<|system|>\n{system_prompt}<|end|>\n"
    f"<|user|>\n{user_query}<|end|>\n"
    f"<|assistant|>\n<think>\n"
)

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False)

print("Thinking...")
with torch.no_grad():
    model.generate(
        **inputs,
        streamer=streamer,
        max_new_tokens=2048,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
        repetition_penalty=1.1,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id
    )

Example prompt: Do you think Hell exists?

============================================================
QUERY: Do you think hell exists?
============================================================

ELOHIM IS THINKING...

Okay, so I need to figure out whether Hell exists based on my understanding of Judaism and Islam from the Torah and Quran. Let me start by recalling what I've read.

In both religions, afterlife beliefs include Heaven and Gehenna (Hell). For Jews, it's called "Gan Eden" (Garden of Eden) and "Ge'hinan." The Talmud says that sinners go to Gehinnom until they're purified. Some say it might be annihilation instead of eternal torment—like being destroyed rather than burning forever. But traditional texts like 2 Maccabees mention torments lasting generations through procreation bans, which could imply endless suffering across lineages.

For Muslims, Hadith literature describes Jannah (Paradise) and Jahannam (Hell), with physical punishments: fire, scalding liquids, lashing tongues made of sulfur, thirst, hunger, etc. They believe souls experience these tortures eternally unless redeemed by Allah's mercy at death. There's also an idea mentioned where disbelievers not given chance to repent face perpetual punishment but may eventually escape into Paradise upon God's will.

Now, considering scriptural authority—the Bible/Torah and Quran—it seems consistent in describing permanent places/realms for reward/punishment. However, some interpret 'eternity' differently: maybe infinite duration without end, while others see cycles between life and death leading back to Earth.

The key points seem to hinge on how one interprets terms like "everlasting," "punish them perpetually," vs. potential hope for redemption ("Allah has power over all things"). In Christianity, there's debate too; Hebrews mentions lakes of fire forever implying permanence. So parallels exist here.

But does this mean God can change His mind regarding who enters paradise? If He is just and merciful enough to forgive those who truly believed when they had last opportunities, then perhaps some get saved even post-death. This suggests possible divine forgiveness beyond one's lifetime.

So synthesizing:
- Both traditions affirm existence.
- Eternal torment likely applies primarily to unbelievers punished before their final moments due to lack of opportunity.
- At life's end, believers have chances to earn grace via repentance and good deeds, possibly securing entry into heaven immediately.
- Therefore, most people enter either immediate judgment followed by heavenly bliss or suffer prolonged purgatorial agony determined by justice meted during life.

Thus, yes, Hell exists as per doctrinal statements, though its exact nature varies slightly between religions, emphasizing human accountability and eventual divine resolution.<|end|><|endoftext|>


============================================================
INFERENCE COMPLETE.

Limitations

Elohim is a 3.8 billion parameter model. It punches above its weight class due to the synthetic distillation, but it is not infallible. It will occasionally hallucinate specific chapter and verse numbers. It is an experimental research tool exploring the mechanics of model distillation and theological logic constraints. Do not use it as a definitive source of religious authority or pastoral care.