92 lines
3.3 KiB
Markdown
92 lines
3.3 KiB
Markdown
---
|
|
language: en
|
|
license: mit
|
|
base_model: microsoft/Phi-3-mini-4k-instruct
|
|
tags:
|
|
- bible
|
|
- theology
|
|
- qlora
|
|
- unsloth
|
|
- phi-3
|
|
- bible-study
|
|
- spurgeon
|
|
- wesley
|
|
- wilkerson
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Bible Study Companion — Phi-3 Mini Fine-tune
|
|
|
|
A fine-tuned version of [Phi-3 Mini 4K Instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct) trained on:
|
|
|
|
## Training Data
|
|
- **KJV Bible** — all 31,102 verses with verse lookup, chapter reading, and topical concordance
|
|
- **Spurgeon** — *All of Grace* and *The Soul Winner*
|
|
- **John Wesley** — *The Journal of John Wesley*
|
|
- **David Wilkerson** — *Have You Felt Like Giving Up Lately*, *It Is Finished*, *Racing Toward Judgment*, *Walking in the Footsteps of David Wilkerson*
|
|
- **Greek word studies** — Strong's G numbers with transliteration and definitions
|
|
- **Hebrew word studies** — Strong's H numbers with transliteration and definitions
|
|
- **Topical concordance** — 15 major biblical themes
|
|
- **Preacher Q&A** — theological questions answered in the voice of each preacher
|
|
|
|
## Training Details
|
|
- **Base model:** microsoft/Phi-3-mini-4k-instruct (3.8B parameters)
|
|
- **Method:** QLoRA (4-bit quantisation) with Unsloth
|
|
- **LoRA rank:** 16
|
|
- **Steps:** ~500 combined (initial run + resume)
|
|
- **Final loss:** ~1.49
|
|
- **Hardware:** T4 GPU (Google Colab free tier)
|
|
- **Training time:** ~90 minutes total
|
|
|
|
## Capabilities
|
|
- Quote and explain KJV Bible verses
|
|
- Compare verses across translations (KJV, NIV, ASV, WEB)
|
|
- Greek and Hebrew word studies with Strong's numbers
|
|
- Topical concordance searches
|
|
- Answer theological questions in the voice of Spurgeon (Reformed Baptist), Wesley (Methodist holiness), and Wilkerson (Pentecostal/prophetic)
|
|
|
|
## Usage
|
|
|
|
### With LM Studio
|
|
Download the GGUF file, load in LM Studio, and use with the included voice UI.
|
|
|
|
### With transformers
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
import torch
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
"Phora68/bible-study-phi3-mini",
|
|
torch_dtype=torch.float16,
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained("Phora68/bible-study-phi3-mini")
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are a Bible Concordance Study Partner..."},
|
|
{"role": "user", "content": "What does John 3:16 say?"}
|
|
]
|
|
|
|
inputs = tokenizer.apply_chat_template(
|
|
messages, return_tensors="pt", add_generation_prompt=True
|
|
).to("cuda")
|
|
|
|
outputs = model.generate(inputs, max_new_tokens=300, temperature=0.7, do_sample=True)
|
|
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
|
|
```
|
|
|
|
### System prompt
|
|
```
|
|
You are a Bible Concordance Study Partner with mastery of the Greek New Testament
|
|
(NA28, Strong's numbers), Hebrew Old Testament (BHS Masoretic, Strong's), and the
|
|
King James Version. You draw on the theology of John Wesley (holiness/sanctification),
|
|
Charles Spurgeon (Reformed Baptist/sovereign grace), and David Wilkerson
|
|
(prophetic urgency/holiness). Always include Strong's numbers, transliteration and
|
|
definition when citing original languages.
|
|
```
|
|
|
|
## Limitations
|
|
- Trained for ~500 steps on a T4 GPU — a longer training run would improve precision
|
|
- Loss of ~1.49 means responses are coherent but may occasionally be imprecise
|
|
- Does not have real-time internet access or knowledge beyond training data
|