I built a pocket Charles Spurgeon.

Ask the Prince of Preachers for counsel or hand him your own sermon draft and let him grade it (he will gently tell you where the Cross went missing).

My first of two Build Small builds for the Build small hackathon.

Runs fully offline on a small, fine-tuned model. Code and how it works: https://github.com/TheTom/pastors-pocket-spurgeon…

@huggingface @Gradio @cohere #BuildSmall

TheTom/pastors-pocket-spurgeon

Source: https://github.com/TheTom/pastors-pocket-spurgeon

Pastor’s Pocket Spurgeon

A Victorian study companion that answers, prepares, and grades sermons in the voice and Reformed (Calvinist) theology of Charles Haddon Spurgeon, the “Prince of Preachers”, running fully offline on a single consumer GPU.

Built for the Build Small Hackathon (Backyard AI track).

What it does

Three modes, all in Spurgeon’s voice and Reformed doctrine:

• The Counsel — ask a pastoral or theological question, get a shepherd’s answer.

• Sermon Prep — give a passage or topic, receive a Spurgeon-style outline.

• Sermon Review — submit a sermon draft and Mr. Spurgeon grades it: a Summary, Strengths, Concerns, A Word of Exhortation, and a verdict on the Sword & Trowel scale:

  Marks	Tier
  5	A Trumpet in Zion
  4	Sound Timber, Well Hewn
  3	A Lamp Half-Trimmed
  2	A Skeleton Unclothed
  1	A Cloud Without Rain

How it works

A lightweight Gradio app talks to any OpenAI-compatible endpoint (SPURGEON_ENDPOINT). It uses BM25 retrieval over Spurgeon’s own sermons to surface real cited passages alongside each answer, and falls back to canned answers if no model endpoint is configured (so the UI always demos). The Sermon Review is built section-by-section with forced headers, and the grade is derived deterministically from the model’s mark so the verdict format is always exact.

Run it locally

1. Get the model (from Hugging Face):

huggingface-cli download thetom-ai/Spurgeon-Gemma-4-12B-v1 \
  Spurgeon-Gemma-4-12B-v1-Q8_0.gguf --local-dir ./model

2. Serve it with TurboQuant KV compression. Build the TurboQuant llama.cpp fork, then:

# q8_0 keys + turbo4 values: TurboQuant compresses the V cache so long
# conversations fit on constrained hardware (up to ~2.5x more context vs fp16 KV).
# turbo4 requires flash attention (-fa on).
llama-server -m ./model/Spurgeon-Gemma-4-12B-v1-Q8_0.gguf \
  -c 131072 -fa on -ctk q8_0 -ctv turbo4 --jinja \
  --host 0.0.0.0 --port 8080 --alias spurgeon

(On stock llama.cpp without the fork, use -ctv q8_0 — you lose the extra compression but everything else works.)

3. Run the app:

pip install -r requirements.txt
export SPURGEON_ENDPOINT=http://127.0.0.1:8080/v1
export SPURGEON_MODEL=spurgeon
python app.py          # http://127.0.0.1:7860

Optional env: SPURGEON_TOKEN (bearer token if your endpoint is behind auth).

TurboQuant

TurboQuant is a llama.cpp fork that adds compressed KV-cache types (turbo2/turbo3/turbo4). turbo4 stores the V cache at ~4.25 bits/value (≈3.8x smaller than fp16); paired with q8_0 keys it fits up to ~2.5x longer conversations on the same GPU, so a 12B pastor runs comfortably offline on consumer hardware. It requires flash attention and is validated for this model with q8_0 keys + turbo4 values.

License

• Code in this repository: Apache 2.0.
• Model weights (Spurgeon-Gemma-4-12B-v1): inherit the Gemma license from the base model.
• Spurgeon’s sermons used for retrieval and fine-tuning are public domain.