ZaraEmbeddings

Standard Works contextual embeddings from ZaraEmbed. This dataset holds 41,995 verses from the Standard Works of The Church of Jesus Christ of Latter-day Saints. Each verse comes with a generated chapter-context block and a 1024-dimension ZaraEmbed embedding of the context and verse together. It is a ready-made index for semantic search and retrieval-augmented question answering over the Standard Works, with no embedding API needed.

Trained and prepared by Brigham Young University-Idaho on public-domain text. Chapter headings, footnotes and Church study aids were not used.

ZaraAI family

Part of ZaraAI, BYU-Idaho's scripture-native model collection:

Item	Role
BYU-Idaho/ZaraBERTa	Cased, scripture-native language model
BYU-Idaho/ZaraEmbed	Verse embedder that produced this dataset's vectors
BYU-Idaho/ZaraRerank	Cross-encoder reranker for ZaraEmbed's top 100

Purpose

These models and data are released for research on scripture and on domain-specific language models. They are also released as aids for personal scripture study. They are offered in the hope that they help readers find and understand the Book of Mormon, and come unto Christ.

What is in it

Column	Type	Description
id	string	Stable identifier, for example book-of-mormon_1-nephi_3_7
reference	string	Human-readable citation, for example 1 Nephi 3:7
volume	string	One of the five volumes (distribution below)
book	string	Book name, for example 1 Nephi, Isaiah, Doctrine and Covenants
chapter	int32	Chapter or section number
verse	int32	Verse number
verse_text	string	The verse alone, with no additions
chapter_context	string	Generated summary of the chapter this verse belongs to
contextual_text	string	The context and verse block that was embedded
embedding	list[float32] (1024)	ZaraEmbed over contextual_text, mean pooled and L2-normalised

Volume	Verses
Old Testament	23,145
New Testament	7,957
Book of Mormon	6,604
Doctrine and Covenants	3,654
Pearl of Great Price	635
Total	41,995

Usage

Embed questions with ZaraEmbed the same way the verses were embedded: prefix a single space, mean-pool last_hidden_state over real tokens, then L2-normalise. Questions use up to 160 tokens. Because the vectors are normalised, cosine similarity is a dot product.

import numpy as np, torch, torch.nn.functional as F
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModel

ds = load_dataset("BYU-Idaho/ZaraEmbeddings", split="train")
V = np.asarray(ds["embedding"], dtype=np.float32)            # (41995, 1024)

tok = AutoTokenizer.from_pretrained("BYU-Idaho/ZaraEmbed")
model = AutoModel.from_pretrained("BYU-Idaho/ZaraEmbed", add_pooling_layer=False).eval()

def embed(q):
    batch = tok([" " + q], truncation=True, max_length=160, return_tensors="pt")
    with torch.no_grad():
        h = model(**batch).last_hidden_state
    m = batch["attention_mask"].unsqueeze(-1).to(h.dtype)
    return F.normalize((h * m).sum(1) / m.sum(1), dim=-1)[0].numpy()

scores = V @ embed("Who were the people Mosiah found living in Zarahemla?")
for i in np.argsort(-scores)[:5]:
    print(ds[int(i)]["reference"], round(float(scores[i]), 3))

To rerank, pass the top 100 verses' verse_text to ZaraRerank. The reranker was trained on bare verse text, not on contextual_text.

How the contextual vectors compare

Right verse first (top-1). The no-names test holds 289 questions on 37 Book of Mormon, D&C and Pearl of Great Price chapters that no ZaraEmbed training pair touched, with no names or distinctive words from the verse. The human benchmark holds 152 human-written study questions across all five volumes.

Test	ZaraEmbed, bare verses	ZaraEmbed, contextual (this dataset)	OpenAI, contextual
No-names (289)	191	203	94
Human benchmark (152)	92	94	120
Human benchmark, Restoration questions (63)	36	37	41

Context helps ZaraEmbed a little on both tests. On the human benchmark, OpenAI's contextual embeddings still lead, mostly on Bible questions. The story-style held-out test is left out of this table because its questions were written from the same chapter summaries that contextual_text contains, which would inflate the contextual scores.

Limitations and responsible use

• These are study aids, not doctrinal authorities, and they are not an official publication of The Church of Jesus Christ of Latter-day Saints.
• Retrieval scores measure textual relevance, not interpretation; read every verse in its context.
• The chapter summaries are model-generated and interpretive. Do not quote them as scripture or as commentary.
• Every verse in a chapter shares one summary, so verses within a chapter are harder to tell apart. ZaraRerank, which reads bare verse text, helps separate them.
• Results come from a single training seed, and the held-out questions were written by a language model.

Provenance and license

• Verse text comes from public-domain editions. Chapter headings, footnotes, cross-references and study aids were excluded.
• chapter_context is generated summary text created for the source dataset, not taken from any published commentary.
• Embeddings were produced with ZaraEmbed over contextual_text at up to 256 tokens. Four rows were re-embedded on CPU as a check and matched the stored vectors at cosine 0.9999 or higher.

Trained and evaluated by Brigham Young University-Idaho, 2026. Released under CC BY 4.0. Verse text is public domain; the license covers the chapter summaries, embeddings and dataset arrangement. ZaraEmbed itself is MIT licensed.

Citation

@misc{zaraembeddings_2026,
  title  = {ZaraEmbeddings: Standard Works contextual embeddings from ZaraEmbed},
  author = {Vallejo, Ron and Brigham Young University-Idaho},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BYU-Idaho/ZaraEmbeddings}
}