Rhema AI

We build open-source AI tools for Bible study.

Rhema is a Bible study platform: reader, commentary engine, learning system. While building it we kept hitting the same wall with general-purpose language models. They paraphrase when they should quote. They flatten theological nuance into mush. They make up verse references. Ask about a contested doctrine and you get one perspective dressed up as consensus.

So we started training our own.

What we're working on

We're fine-tuning Bible-specialized models from open-weight bases (Qwen 2.5 7B right now). Two-stage pipeline: continued pretraining on public-domain theological corpora (Calvin, Barnes, Pulpit Commentary, Keil-Delitzsch, creeds and confessions), then QLoRA instruction tuning on 50,000+ supervised examples.

Those examples come from 24 synthetic data generators we wrote, each targeting a different slice of biblical scholarship. Verse lookup, passage exposition, Hebrew and Greek exegesis, cross-references, doctrinal Q&A, patristic readings, creedal analysis, multi-tradition comparison. All of it grounded in the Berean Standard Bible with Hebrew morphology, Greek lexicon data, and Strong's numbers across the full 31,102-verse canon.

What the model actually does

It quotes scripture exactly from the BSB. No paraphrasing, no "the Bible says something like..." If it's interpretation, the model says so.

It handles theological disagreements by presenting Protestant, Catholic, and Orthodox positions without picking a side. Hebrew and Greek terms show up with transliteration, parsing, and Strong's numbers because the original languages carry meaning that English translations compress.

It stays within theology, biblical history, and Christian thought. It doesn't try to be a counselor.

Under the hood

Training produces LoRA adapters that merge back to the base and export to GGUF for local inference via Ollama / llama.cpp. Quantized from Q3_K_M (6.5 GB) to Q5_K_M (10 GB) so they run on consumer GPUs.

Retrieval: ChromaDB vector store with BGE embeddings over the full BSB, plus SQLite databases for verse search (FTS5), morphology, lexicon, and cross-references. Training on RunPod (A100 / A40).

Connection to Rhema

These models power features inside Rhema: the Study Canvas AI assistant, commentary suggestions, passage analysis. The platform is where the models get tested against real questions, and where the gaps show up. That's what makes the next round of training data better.

What's next

We're working toward publishing our first models and datasets on this org. If you're building Bible study tools or doing anything with theological NLP, get in touch: hello@rhemabible.co