---
library_name: transformers
language:
- en
tags:
- bert
- domain-adaptation
- religious-text
- biblical-corpus
- masked-language-modeling
- nlp
- text-classification
- question-answering
- sentence-similarity
license: apache-2.0
base_model: bert-base-uncased
pipeline_tag: fill-mask
---

# ReligionBERT

ReligionBERT is a domain-adapted BERT model produced by continued masked language modelling (MLM) pre-training on the English Bible corpus. Starting from `bert-base-uncased`, the model was trained for 30,000 steps on 62,197 verses drawn from the King James Version and World English Bible translations. It is designed for downstream NLP tasks involving religious and biblical text, where general-purpose BERT models underperform due to archaic language, theological vocabulary, and long-range intertextual dependencies.

A companion multilingual model, [MultiReligionBERT](https://huggingface.co/LucasLicht/multi-religion-bert), covers 12 languages and supports cross-lingual zero-shot transfer for African language religious NLP.

---

## Model Details

| Field | Details |
|---|---|
| **Model type** | BERT (encoder-only, masked language model) |
| **Base model** | `bert-base-uncased` (109M parameters) |
| **Pre-training objective** | Continued MLM (15% token masking) |
| **Pre-training corpus** | English Bible (KJV + WEB), 62,197 verses |
| **Training steps** | 30,000 |
| **Final validation loss** | 1.164 |
| **Language** | English |
| **License** | Apache 2.0 |
| **Developed by** | Lucas Licht |
| **Institution** | Koforidua Technical University, Ghana |

---

## Intended Use

ReligionBERT is intended for NLP tasks on religious and biblical text, including:

- **Semantic similarity** between Bible verses or passages
- **Book and section classification** of biblical text
- **Extractive question answering** over religious passages
- **Feature extraction** for downstream religious NLP pipelines
- **Masked language modelling** for biblical text generation and completion

It is not recommended for general-domain NLP tasks where `bert-base-uncased` is likely a stronger baseline.

---

## How to Get Started

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch

tokenizer = AutoTokenizer.from_pretrained("LucasLicht/religion-bert")
model = AutoModelForMaskedLM.from_pretrained("LucasLicht/religion-bert")

text = "For God so loved the [MASK] that he gave his only begotten Son."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

masked_index = (inputs["input_ids"] == tokenizer.mask_token_id).nonzero(as_tuple=True)[1]
logits = outputs.logits[0, masked_index]
predicted_token = tokenizer.decode(torch.argmax(logits, dim=-1))
print(predicted_token)  # "world"
```

For sentence embeddings or downstream fine-tuning:

```python
from transformers import AutoTokenizer, AutoModel
import torch

tokenizer = AutoTokenizer.from_pretrained("LucasLicht/religion-bert")
model = AutoModel.from_pretrained("LucasLicht/religion-bert")

inputs = tokenizer("The Lord is my shepherd; I shall not want.", return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)

# Use CLS token as sentence representation
cls_embedding = outputs.last_hidden_state[:, 0, :]
```

---

## Training Details

### Pre-Training Corpus

The corpus was sourced from the [christos-c/bible-corpus](https://github.com/christos-c/bible-corpus) repository. All 66 books of the King James Version and World English Bible translations were extracted via XML parsing, yielding 62,197 verses.

### Training Procedure

| Hyperparameter | Value |
|---|---|
| Base model | `bert-base-uncased` |
| Training steps | 30,000 |
| Effective batch size | 32 (16 per device, 2 gradient accumulation steps) |
| Learning rate | 3e-5 (linear warmup, 500 steps) |
| Weight decay | 0.01 (AdamW) |
| MLM masking probability | 15% |
| Max sequence length | 128 tokens |
| Precision | FP16 mixed precision |
| Hardware | NVIDIA Tesla T4 / A100 (Google Colab) |
| Framework | HuggingFace Transformers 5.0.0 |

Training was conducted across multiple sessions with checkpoint recovery. A `PermanentDeleteCallback` retained only the two most recent checkpoints to prevent storage exhaustion. All metrics were logged to Weights and Biases.

### Training Loss Curve

| Step | Validation Loss |
|---|---|
| 500 | 1.806 |
| 5,000 | 1.451 |
| 10,000 | 1.339 |
| 15,000 | 1.272 |
| 20,000 | 1.204 |
| 25,000 | 1.156 |
| 28,500 | **1.129** (best) |
| 30,000 | 1.164 (final) |

---

## Evaluation

### Perplexity on Held-out Religious Text

Perplexity was computed on 500 held-out English Bible verses not seen during pre-training.

| Model | Perplexity (lower is better) |
|---|---|
| `bert-base-uncased` | 15.48 |
| **ReligionBERT** | **3.73** |

ReligionBERT achieves a **75.9% reduction** in perplexity, confirming successful domain alignment.

---

### Downstream Task Results

Three fine-tuning tasks were evaluated using three automatically constructed datasets. All results are on held-out test sets. ReligionBERT is compared to its generic baseline (`bert-base-uncased`).

#### Semantic Similarity (21,994 verse pairs)

| Model | Pearson | Spearman |
|---|---|---|
| bert-base-uncased | 0.9569 | 0.6556 |
| **ReligionBERT** | **0.9623** | **0.6591** |

#### Book Classification (7,726 samples, 66 classes)

| Model | Accuracy | Macro F1 |
|---|---|---|
| bert-base-uncased | 0.4075 | 0.3144 |
| **ReligionBERT** | **0.4347** | **0.3381** |

#### Extractive Question Answering (1,199 LLM-assisted examples)

| Model | Exact Match (%) | Token F1 (%) |
|---|---|---|
| bert-base-uncased | 32.50 | 58.91 |
| **ReligionBERT** | **40.00** | **60.48** |

ReligionBERT outperforms `bert-base-uncased` on **5 of 6 metrics**, with the strongest gain of **+7.50 Exact Match points** on extractive QA.

---

## Datasets

The three fine-tuning datasets used in this study are described below. They are derived automatically from the Bible corpus without manual annotation (except for the QA human verification step).

| Dataset | Task | Size | Notes |
|---|---|---|---|
| Verse Similarity | Semantic similarity (STS) | 21,994 pairs | Cross-translation and intra-corpus pairs; balanced subset 6,392 |
| Bible Book Classification | Text classification | 7,726 samples | 66 classes (all Bible books); 80/10/10 split |
| Bible QA | Extractive QA | 1,199 examples | LLM-assisted via Llama 3.3 70B; SQuAD v2 format; 100% human-verified quality |

---

## Limitations

- Pre-training is limited to the English Bible (KJV and WEB translations). Performance on other religious traditions (Quran, Vedas, Buddhist sutras) has not been evaluated.
- The model inherits biases present in `bert-base-uncased` and may reflect theological perspectives embedded in the King James Version.
- The QA fine-tuning dataset is relatively small (959 training examples); downstream QA performance may benefit from larger domain-specific QA datasets.
- The model is not intended for general-domain tasks.

---

## Citation

If you use this model, please cite:

```bibtex
@misc{licht2025religionbert,
  title     = {ReligionBERT: Domain-Adaptive Pre-Training of BERT on Biblical Corpora for Religious NLP Tasks},
  author    = {Licht, Lucas},
  year      = {2025},
  note      = {Koforidua Technical University, Ghana. Model available at https://huggingface.co/LucasLicht/religion-bert}
}
```

---

## Contact

For questions or collaboration, reach out via HuggingFace or GitHub: [@Licht005](https://github.com/Licht005)