ainouche-abderahmane/synoptiq-corpus

SynoptiQ Corpus

A token-level dataset of the Synoptic Gospels (Matthew, Mark, Luke) in Koine Greek, annotated with morphological tags, Aland pericope boundaries, and pairwise Needleman-Wunsch token alignments for computational research on the Synoptic Problem.

Overview

The Synoptic Problem --- the question of how Matthew, Mark, and Luke are literarily related --- has occupied New Testament scholarship for over two centuries. This dataset provides a machine-learning-ready representation of the synoptic gospels designed for training transformer-based models to detect copying direction, editorial tendencies, and source relationships.

Each row is a single token with its lemma, part-of-speech tag, morphological parsing, and Aland pericope assignment. Parallel passages across gospels are aligned at the token level using Needleman-Wunsch global alignment keyed on (lemma, POS) pairs, enabling models to learn from the correspondence structure directly.

Key properties

• 49,061 tokens across Matthew (18,329), Mark (11,286), and Luke (19,446)
• 170 Aland pericopes with tradition classification: 88 triple (all three synoptics), 17 double (Matthew--Luke only), 18 Matthean-only, 45 Lukan-only, 2 Markan-only
• 235 pairwise token alignments (Matthew-Mark, Matthew-Luke, Mark-Luke) computed per shared pericope, yielding 8,855 aligned token pairs and 45,029 gap positions
• 2,739 unique lemmas with MorphGNT morphological annotations (part-of-speech, tense, voice, mood, case, number, gender)
• Stratified 60/20/20 split at the pericope level: 101 train, 33 validation, 36 test. No verse appears in more than one split
• SBLGNT text, the standard critical edition published by the Society of Biblical Literature (CC-BY), merged with MorphGNT annotations (CC-BY-SA)

Data sources

Source	Description	License
SBLGNT	Society of Biblical Literature Greek New Testament (Holmes 2010)	CC-BY
MorphGNT	Morphologically annotated SBLGNT (Tauber 2017)	CC-BY-SA
Aland Synopsis Quattuor Evangeliorum	Pericope numbering and verse mapping (Aland 1963)	Academic reference

Limitations

• Covers the synoptic gospels only; John is excluded as it is not directly relevant to the Synoptic Problem.
• Pericope assignment covers 96% of tokens; approximately 1,984 tokens (4%) fall outside the Aland table and carry an empty pericope_id.
• Morphological annotations reflect one scholarly analysis (MorphGNT); alternative parsings exist but are not included.
• The Needleman-Wunsch alignment uses lemma+POS matching with fixed gap penalties; different scoring parameters may produce slightly different alignments.

Intended Uses

• Training and evaluating NLP models for Koine Greek (POS tagging, lemmatisation, morphological analysis)
• Textual criticism: detecting scribal tendencies, editorial fatigue, and patterns of textual dependence
• Stylometric analysis and authorship studies of New Testament texts
• Pedagogical tools for Koine Greek instruction and self-study
• Benchmarking domain adaptation techniques for historical languages

Related Models

• KoineFormer — LoRA-adapted GreTa model trained on Koine Greek, evaluated on this corpus.

Dataset structure

Main table: token-level data (data/*.parquet)

Each row is one token. The dataset is split into train, validation, and test Parquet files.

Column	Type	Description
token_id	string	Stable identifier: Matthew.1.1.0
book	string	Canonical gospel name: Matthew, Mark, Luke
chapter	int32	Chapter number (1-indexed)
verse	int32	Verse number (1-indexed)
position	int32	Token position within verse (0-indexed)
text	string	Surface form as written in SBLGNT: γενέσεως
normalized	string	NFD-normalised, accent-stripped: γενεσεως
lemma	string	Dictionary headword from MorphGNT: γένεσις
pos	string	Part-of-speech tag: N- (noun), V- (verb), RA (article), C- (conjunction), RP (relative pronoun), P- (preposition), A- (adjective), D- (adverb)
morph	string	8-character CCAT morphological tag encoding person, tense, voice, mood, case, number, gender, degree
pericope_id	string	Aland pericope number: 018; empty string if unassigned
tradition	string	triple, double, matthean_unique, lukan_unique, mark_unique
genre	string	narrative, discourse, wisdom, passion, other
books_in_pericope	string	JSON array: '["Matthew","Mark","Luke"]'
is_punctuation	bool	Always false in SBLGNT; present for schema compatibility

Auxiliary files

• pericopes.parquet --- Pericope-level metadata (170 rows): pericope_id, tradition, genre, books, n_tokens, n_alignment_pairs, split. Useful for stratified sampling and per-pericope analysis without joining the token table.
• alignments.json --- Token-level Needleman-Wunsch alignment pairs, keyed by pericope_id|book_a|book_b. Each value is a list of [idx_a, idx_b] pairs where null indicates a gap. Indices reference the position of the token within the pericope's per-book token sequence (sorted by chapter, verse, position).

Usage

from datasets import load_dataset

# Load the token-level dataset with train/val/test splits
dataset = load_dataset("ainouche-abderahmane/synoptiq-corpus")

# Filter to triple tradition only
triple = dataset["train"].filter(
    lambda x: x["tradition"] == "triple"
)

# Extract all lemmas from a specific book
matthew_lemmas = dataset["train"].filter(
    lambda x: x["book"] == "Matthew"
)["lemma"]

For fine-tuning a masked language model on Koine Greek:

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("ainouche-abderahmane/synoptiq-corpus")
tokenizer = AutoTokenizer.from_pretrained("bowphs/GreTa")

def tokenize(batch):
    return tokenizer(batch["text"], truncation=True, max_length=128)

tokenized = dataset.map(tokenize, batched=True)

Alignment format

The alignments.json file uses the key format pericope_id|book_a|book_b:

{
  "018|Matthew|Mark": [
    [0, null],
    [1, 0],
    [2, null],
    [null, 1]
  ]
}

• [0, null] --- Matthew token 0 has no counterpart in Mark (gap in Mark)
• [1, 0] --- Matthew token 1 aligns with Mark token 0
• [null, 1] --- Mark token 1 has no counterpart in Matthew (gap in Matthew)

Indices reference the per-book token sequence for that pericope, ordered by (chapter, verse, position). Use the token table to join indices back to surface forms and lemmas.

Citation

If you use this dataset, please cite:

@dataset{synoptiq_corpus,
  title     = {SynoptiQ Corpus: A Token-Aligned Dataset of the Synoptic
               Gospels for Computational Source Criticism},
  author    = {Ainouche, Abderahmane},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/ainouche-abderahmane/synoptiq-corpus}
}

The underlying text and annotations were created by:

• Holmes, M. W. (2010). The Greek New Testament: SBL Edition. Society of Biblical Literature. CC-BY.
• Tauber, J. K. (2017). MorphGNT: Morphologically Parsed Greek New Testament. CC-BY-SA.
• Aland, K. (1963). Synopsis Quattuor Evangeliorum. Deutsche Bibelgesellschaft.

License

This dataset is licensed under CC-BY-SA 4.0, in accordance with the share-alike requirement of MorphGNT.

Version history

Version	Date	Changes
1.0.0	2026-07	Initial release. SBLGNT + MorphGNT merge, Aland pericope assignment, Needleman-Wunsch alignments, stratified splits.