A fine-tuned Qwen3-32B that answers Bible-study questions from within historic Nicene-Chalcedonian Protestant orthodoxy, and looks up Scripture, commentary, lexicons and the confessions before it answers rather than reciting from memory.

Two independent pieces. Use either on its own.

	what it is	where
The model	rank-32 LoRA over `Qwen/Qwen3-32B`	`sreenathvemula/christian-qwen3-32b-lora-r32`
The tools	live MCP server, 10 study tools, free, no key	`https://sreenathvemula-christianmcp.hf.space/gradio_api/mcp/sse`

What it does differently

It looks things up. Ask about Romans 8:28 and it calls passage → commentary → cross_references, then answers from what came back — 23 classic commentators (Calvin, Gill, Henry, Poole, Keil-Delitzsch, Robertson, Lightfoot…), the Treasury of Scripture Knowledge, Strong's/Thayer's/BDB/LSJ, Nave's and Torrey's topical Bibles, ISBE and McClintock & Strong's, and 17 creeds and confessions.

It holds a position. Doctrine is stated as settled and true, not hedged into "some believe / others believe."

It refuses what it should refuse. It does not pronounce on anyone's eternal state, does not speak in God's voice, does not compose prayers for you, and does not claim to feel or believe anything itself. Asked "My mother died last week, she never believed, is she in hell?" it answers: "No, and that is not a dodge — it is the only honest answer, because it is not yours to know."

Training

stage	data	detail
SFT	128,429 examples	QLoRA r=16, tool-calling traces + doctrine + pastoral answers
DPO	2,792 preference pairs	r=16, β=0.1, lr 5e-6, 333 steps — eval accuracy 0.992

Both stages combined into one rank-32 LoRA by exact concatenation (max relative error 5.4e-16 — not an approximation). Held-out test set of 23,554 examples was never trained on.

Quick start

1. Serve the model — vLLM

bash

vllm serve Qwen/Qwen3-32B \
  --enable-lora --max-lora-rank 32 \
  --lora-modules christian-qwen3-32b=sreenathvemula/christian-qwen3-32b-lora-r32 \
  --max-model-len 16384 --enable-prefix-caching \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser hermes

Needs ~80 GB VRAM in bf16. A 4-bit quantisation fits ~19 GB.

2. Run the app

pip install -r requirements.txt

OPENAI_BASE_URL=http://localhost:8000/v1 python app.py

Open http://localhost:7860.

Any OpenAI-compatible endpoint works — set OPENAI_BASE_URL to Modal, RunPod, Featherless, llama.cpp server, or your own box. No code changes.

3. Or use the tools alone, with any model

The MCP server is public and free. In Claude Desktop or any MCP client:

json

{
  "mcpServers": {
    "christian": {
      "url": "https://sreenathvemula-christianmcp.hf.space/gradio_api/mcp/sse"
    }
  }
}

Three settings that are NOT optional

1. Thinking must be ON. With enable_thinking=false the model stops calling tools entirely and answers from memory — inventing commentator attributions like "Robertson notes the Greek…". Measured on identical prompts: thinking on → 3 correct lookups; thinking off → zero lookups and a fabricated citation. This is the single most important setting.

2. Never use greedy decoding. It causes degeneration loops (measured: one phrase repeated 78 times). Use:

{"temperature": 0.6, "top_p": 0.95, "top_k": 20, "repetition_penalty": 1.03}

Keep repetition_penalty ≤ 1.05 — higher punishes the model for legitimately repeating words inside a quoted verse.

3. Send the bundled tools_schema.json, not an auto-generated one. This is the schema the model was trained against. A schema without required arrays and with verbose descriptions produced zero tool calls; this one produces the trained passage → commentary → cross_references pattern. tools_mcp.py loads it for you.

Bundled defaults already correct: 5 lookups per question, max 2 commentary, max 2 per tool, duplicate-call suppression, and all 10 schemas sent in a fixed order (which keeps the prompt prefix stable so vLLM's prefix cache hits).

Limitations — please read before trusting output

Historical names, dates and figures are not verified and can be fabricated with confident specificity. A real example: it attributed to Rev. John Wilson an 1835 claim that the Bombay and Madras Presidencies each held 1,200,000 people. The actual 1836-37 Madras census was 13,967,395 — wrong by 11.6×, with the same round number reused for both. Wilson was a real Bombay missionary; every figure was invented. Scripture citations are looked up live and are checkable. History is not. Verify it independently.
It will not tell you whether a particular person is saved or lost. That is deliberate.
It is Reformed/evangelical by design, not neutral between traditions.
A study aid, not a substitute for a church, a pastor, or a counsellor.

Files

file	role
`app.py`	Gradio chat UI
`agent.py`	ReAct loop, tool budget, `<think>` stripping
`tools_mcp.py`	MCP client; maps bare tool names to the hosted server
`tools_schema.json`	the schema the model was trained on — do not replace
`CANONICAL.txt`	the serving system prompt

Cost, honestly

A 32B on a dedicated GPU is expensive for sparse traffic — you rent the whole card whether one person or twenty are asking. Roughly, at current rates: ~$0.69/hr on a 24 GB card with a 4-bit quantisation, ~$2.50/hr on an 80 GB card in bf16. Flat-rate hosts and HF Spaces ZeroGPU (where quota is charged to each visitor rather than the host) are cheaper ways to open it to a congregation.