makarios19/IndoBerea-v1

IndoBerea v1

Rebuilt model for IndoBerea: Evolving Semantic Search in Theological Context (F. V. P. Samosir and S. Mendrofa, ICIC 2023, IEEE): semantic search over the Indonesian Terjemahan Baru (TB) Bible.

• Paper (IEEE Xplore): https://ieeexplore.ieee.org/document/10382053/
• Results page (static): https://huggingface.co/spaces/makarios19/IndoBerea-v1
• Code, known issue, and roadmap: https://github.com/FeliksMakarios/IndoBerea

Limitations at a glance. This model is effectively untrained IndoBERT with mean pooling (see below). It is published to document and reproduce v1, not for Bible study. There is no live demo: the Space is a static page with the precomputed results for the seven queries in the paper. IndoBerea v2 is in development.

What this model is

indobenchmark/indobert-base-p2 followed by mean pooling, loaded with SentenceTransformers.

The original weights from 2023 were lost. They do not need to be recovered, because the original training run produced no updates: the notebook built InputExample(texts=[verse]) without labels and trained with BatchHardTripletLoss. With one label for every example, that loss is constant and its gradient is zero (a runnable check is in the GitHub repository). The only change to the weights during that run came from AdamW weight decay, which shrinks them by well under 0.1%.

scripts/build_v1.py in the repository rebuilds this model and compares its top-10 results and cosine scores with the outputs saved in the original notebook. For the seven queries in the paper, the largest score difference is 0.0007 (mean 0.0002), and 66 of 70 verses appear at the same rank; the remaining differences are between verses with nearly tied scores.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("makarios19/IndoBerea-v1")
emb = model.encode(["tetap bersukacita", "Bersukacitalah senantiasa."], normalize_embeddings=True)
print(emb[0] @ emb[1])

Intended use and limitations

• Use this model to reproduce the v1 paper and as a baseline. It is not trained for sentence similarity and is not recommended as a Bible study tool.
• Short verses (for example Yoh 6:48 or 1Te 5:19) appear for many unrelated queries.
• The paper's evaluation is small: seven queries judged by the authors, without inter-annotator agreement or baselines.
• No live demo is provided. The Space shows only precomputed results for the paper's queries.
• No Bible text is distributed with the model; Terjemahan Baru is copyrighted by Lembaga Alkitab Indonesia. To search the Bible, encode a local copy of the text you are licensed to use.
• Results are retrieval aids, not interpretations; always read the surrounding passage.

IndoBerea v2, with a relevance benchmark and properly trained retrievers, is in progress.

Citation

@inproceedings{samosir2023indoberea,
  title     = {IndoBerea: Evolving Semantic Search in Theological Context},
  author    = {Samosir, Feliks Victor Parningotan and Mendrofa, Serius},
  booktitle = {2023 Eighth International Conference on Informatics and Computing (ICIC)},
  publisher = {IEEE},
  pages     = {1--6},
  year      = {2023},
  doi       = {10.1109/ICIC60109.2023.10382053},
  url       = {https://ieeexplore.ieee.org/document/10382053/}
}