LoveJesus/biblical-entity-recognizer-chirho

Biblical Entity Recognizer (Chirho) - Model 6

A DistilBERT-based Named Entity Recognition model fine-tuned on 200K+ annotated tokens from the King James Version (KJV) Bible to recognize six types of biblical entities using BIO tagging.

Repo: LoveJesus/biblical-entity-recognizer-chirho

Model Overview

Property	Value
Base Model	distilbert-base-uncased
Parameters	~66M
Task	Token Classification (NER)
Tagging Scheme	BIO (Beginning-Inside-Outside)
Number of Labels	13
Entity Types	6
Max Sequence Length	128 tokens
Framework	HuggingFace Transformers
License	MIT

Entity Types

Entity Type	BIO Tags	Description	Examples
PERSON	B-PERSON, I-PERSON	Biblical persons and figures	Moses, David, Paul, Mary
DIVINE	B-DIVINE, I-DIVINE	Names and titles of God	God, LORD, Jesus Christ, Holy Spirit
PEOPLE_GROUP	B-PEOPLE_GROUP, I-PEOPLE_GROUP	Nations, tribes, and groups	Israelites, Philistines, Pharisees, Corinthians
PLACE	B-PLACE, I-PLACE	Geographical locations	Jerusalem, Bethlehem, Egypt, Gethsemane
EVENT	B-EVENT, I-EVENT	Biblical events and feasts	Passover, Pentecost, Sabbath
ARTIFACT	B-ARTIFACT, I-ARTIFACT	Sacred objects and instruments	Urim, Thummim

Full Label Set (13 Labels)

O, B-PERSON, I-PERSON, B-DIVINE, I-DIVINE, B-PEOPLE_GROUP, I-PEOPLE_GROUP,
B-PLACE, I-PLACE, B-EVENT, I-EVENT, B-ARTIFACT, I-ARTIFACT

Evaluation Results

Metric	Score
F1	0.9810
Precision	97.78%
Recall	98.43%
Best Epoch	4

Entity-level metrics are computed using the seqeval library, which evaluates complete entity spans rather than individual token labels, providing a rigorous assessment of recognition quality.

Per-Entity F1

Entity Type	F1
DIVINE	0.9993
PLACE	0.9742
PERSON	0.9617
PEOPLE_GROUP	0.9512
Overall (macro)	0.9786

Per-entity metrics evaluated using seqeval on the held-out test set. The model excels at recognizing divine names (F1=0.9993) and performs strongly across all entity types.

Usage

Quick Start: Pipeline API

# For God so loved the world that he gave his only begotten Son,
# that whoever believes in him should not perish but have eternal life. - John 3:16

from transformers import pipeline

ner_pipeline_chirho = pipeline(
    "token-classification",
    model="LoveJesus/biblical-entity-recognizer-chirho",
    aggregation_strategy="simple",
)

text_chirho = "And Moses said unto the LORD in the land of Egypt"
entities_chirho = ner_pipeline_chirho(text_chirho)

for entity_chirho in entities_chirho:
    print(f"{entity_chirho['word']}: {entity_chirho['entity_group']} ({entity_chirho['score']:.3f})")
# Moses: PERSON (0.998)
# LORD: DIVINE (0.999)
# Egypt: PLACE (0.997)

Manual Inference

# For God so loved the world that he gave his only begotten Son,
# that whoever believes in him should not perish but have eternal life. - John 3:16

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer_chirho = AutoTokenizer.from_pretrained("LoveJesus/biblical-entity-recognizer-chirho")
model_chirho = AutoModelForTokenClassification.from_pretrained("LoveJesus/biblical-entity-recognizer-chirho")

text_chirho = "Then Jesus went with them unto a place called Gethsemane."
inputs_chirho = tokenizer_chirho(text_chirho, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs_chirho = model_chirho(**inputs_chirho)
    predictions_chirho = torch.argmax(outputs_chirho.logits, dim=2)

tokens_chirho = tokenizer_chirho.convert_ids_to_tokens(inputs_chirho["input_ids"][0])
pred_ids_chirho = predictions_chirho[0].tolist()
id2label_chirho = model_chirho.config.id2label

for token_chirho, pred_id_chirho in zip(tokens_chirho, pred_ids_chirho):
    label_chirho = id2label_chirho[pred_id_chirho]
    if label_chirho != "O" and token_chirho not in ["[CLS]", "[SEP]", "[PAD]"]:
        print(f"  {token_chirho}: {label_chirho}")
# jesus: B-PERSON
# gethsemane: B-PLACE

Batch Processing

# For God so loved the world that he gave his only begotten Son,
# that whoever believes in him should not perish but have eternal life. - John 3:16

from transformers import pipeline

ner_pipeline_chirho = pipeline(
    "token-classification",
    model="LoveJesus/biblical-entity-recognizer-chirho",
    aggregation_strategy="simple",
)

verses_chirho = [
    "Now when Jesus was born in Bethlehem of Judaea in the days of Herod the king.",
    "And Solomon built the house of the LORD in Jerusalem.",
    "The LORD is my shepherd; I shall not want.",
    "And Paul said unto the Corinthians, Grace be unto you from God our Father.",
]

for verse_chirho in verses_chirho:
    entities_chirho = ner_pipeline_chirho(verse_chirho)
    print(f"\n{verse_chirho}")
    for entity_chirho in entities_chirho:
        print(f"  {entity_chirho['word']}: {entity_chirho['entity_group']} ({entity_chirho['score']:.3f})")

Training Details

Dataset

Property	Value
Source Text	King James Version (KJV) Bible
Text Source	ScrollMapper bible_databases (Public Domain)
Entity Source	STEPBible TIPNR (CC BY) + curated divine names list
Annotated Tokens	200,000+
Annotation Scheme	BIO (Beginning-Inside-Outside)
Format	JSONL with tokens_chirho, ner_tags_chirho, reference_chirho fields
Split Strategy	80/10/10 by book (not verse) to prevent data leakage
Dataset Repo	LoveJesus/biblical-ner-dataset-chirho

Hyperparameters

Parameter	Value
Learning Rate	5.0e-5
Batch Size	32
Epochs	5 (best at epoch 4)
Weight Decay	0.01
Warmup Ratio	0.1
Max Sequence Length	128
Seed	42
Optimizer	AdamW (default Trainer)
Early Stopping Patience	3 epochs
Metric for Best Model	F1 (entity-level)

Subword Alignment

When DistilBERT's WordPiece tokenizer splits a word into multiple subword tokens, only the first subtoken receives the original BIO label. Subsequent subtokens of the same word receive -100 (ignored in loss computation). This prevents the model from being penalized on tokens it cannot meaningfully label.

Hardware Compatibility

The training script supports:

• Apple MPS (Metal Performance Shaders) for Apple Silicon Macs
• CUDA for NVIDIA GPUs
• CPU fallback

Architecture

DistilBERT-base-uncased (66M parameters)
    |
    v
6-layer Transformer Encoder
    |
    v
Token Classification Head (Linear: 768 -> 13)
    |
    v
BIO Label Predictions per Token

DistilBERT is a distilled version of BERT that retains 97% of BERT's language understanding while being 60% faster and 40% smaller. It uses 6 transformer layers (vs. BERT's 12), a hidden size of 768, and 12 attention heads.

Limitations

• Domain Specificity: Trained exclusively on KJV Bible text; may not generalize well to modern English biblical translations or extra-biblical religious texts
• Archaic Language: Optimized for Early Modern English (KJV) vocabulary and syntax ("thou", "unto", "begat")
• Entity Coverage: The six entity categories may not cover all possible biblical entity types (e.g., no separate category for books of the Bible, religious practices, or time periods)
• Base Model Vocabulary: DistilBERT was pre-trained on modern English; some rare biblical proper nouns may be heavily subword-tokenized, potentially reducing recognition accuracy for uncommon names
• Assistive Tool: This model is intended as an assistive tool for Bible study and research, not as a replacement for careful scriptural reading

Intended Use

• Bible study applications requiring automatic entity highlighting
• Biblical text analysis and digital humanities research
• Building knowledge graphs of biblical persons, places, and events
• Enhancing Bible search engines with entity-aware queries
• Educational tools for learning biblical geography, persons, and events

Citation

@misc{lovejesus2026biblicalentityrecognizer,
  title={Biblical Entity Recognizer: DistilBERT NER for KJV Bible Text},
  author={LoveJesus},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/LoveJesus/biblical-entity-recognizer-chirho}
}

License

MIT

---

For God so loved the world that he gave his only begotten Son, that whoever believes in him should not perish but have eternal life. - John 3:16