LoveJesus/biblical-variant-classifier-chirho

Biblical Manuscript Variant Classifier (mT5-small)

Classifies New Testament Greek manuscript variants by type using an mT5-small seq2seq model.

Variant Types

• substitution: Different word forms across critical editions
• omission: Word present in some editions but not others
• addition: Extra words in some editions
• spelling: Minor orthographic variants

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("LoveJesus/biblical-variant-classifier-chirho")
model = AutoModelForSeq2SeqLM.from_pretrained("LoveJesus/biblical-variant-classifier-chirho")

input_text = "classify variant [greek]: βαπτίζω [MAT.3.11] editions: NTS context: ἐγὼ μὲν ὑμᾶς βαπτίζω"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Results (v2)

Trained on Apple M4 Pro (MPS), mT5-small (300M params), 9 epochs:

Metric	Score
Eval loss (best)	0.0868 (epoch 9)
Variant type accuracy	98.87%
Exact match (full output)	37.5%
Training steps	9,558

Per-Epoch Progression

Epoch	Loss	Exact Match	Type Accuracy
1	0.8033	0%	69.3%
2	0.2577	1.0%	67.5%
3	0.1465	12.1%	94.5%
4	0.1159	25.9%	96.0%
5	0.1039	29.3%	98.3%
6	0.0963	32.7%	98.8%
7	0.0914	34.7%	97.8%
8	0.0888	35.6%	98.8%
9	0.0868	37.5%	98.9%

Training Data

Derived from STEPBible TAGNT (Translators Amalgamated Greek NT), which marks each NT word with its presence across 6 critical editions: NA27/28, Textus Receptus, SBLGNT, Byzantine, Westcott-Hort, and THGNT.

• Training set: 8,497 examples
• Validation set: 1,062 examples
• Test set: 1,062 examples

Part of bible.systems

This is model 4 of 8 in the bible.systems ML pipeline.

---

For God so loved the world... — John 3:16