Update dataset card 2026-05-27

README.md

@@ -7,13 +7,17 @@ language:
 - it
 license: cc-by-4.0
 pretty_name: Targum Corpus
+multilinguality:
+- multilingual
 tags:
 - bible
+- bible-translation
 - translation
 - nlp
 - corpus-linguistics
 - parallel-corpus
 - digital-humanities
+- religion
 size_categories:
 - 1M<n<10M
 task_categories:
@@ -21,7 +25,7 @@ task_categories:
 - translation
 configs:
 - config_name: corpora
-  data_files: "corpus/**/*.jsonl"
+  data_files: "corpora/**/*.jsonl"
   default: true
 - config_name: works
   data_files: "works.tsv"
@@ -33,9 +37,9 @@ configs:
   data_files: "book_coverage.tsv"
 ---
 
-# Targum - A Multilingual New Testament Translation Corpus
+# Targum -- A Multilingual New Testament Translation Corpus
 
-![Targum — A Multilingual New Testament Translation Corpus](banner.jpg)
+**Links:** 📄 [LREC 2026 paper](https://aclanthology.org/2026.lrec-main.564/) · 🗂️ [GitHub mirror](https://github.com/mrapacz/targum-corpus) · 🌐 [Online viewer](https://targum.mrapacz.com/)
 
 **Targum** is a multilingual New Testament translation corpus with unprecedented depth in five European languages: English, French, Italian, Polish, and Spanish. It contains **651** translations (**334** unique) collected from **13** source libraries and spanning **1525–2025**.
 
@@ -43,7 +47,13 @@ This dataset contains the **public release subset**: **302** translations distri
 
 Named after the ancient Aramaic translations of the Hebrew Bible (תרגום, "translation"), the corpus prioritizes vertical depth over linguistic breadth, making it possible to computationally analyze a wide spectrum of historical periods and confessional traditions within each language.
 
-Also available on GitHub: [mrapacz/targum-corpus](https://github.com/mrapacz/targum-corpus).
+The repository ships three families of artefacts side by side:
+
+- **Corpus texts** -- verse-level JSONL under `corpora/`.
+- **Embeddings** -- pre-computed encoder vectors under `embeddings/` (~120 GB across six (model, granularity) combinations).
+- **Similarity** -- pairwise lexical and semantic similarity scores under `similarity/` (~5 GB).
+
+The metadata tables (`works`, `editions`, `instances`, `book_coverage`) are exposed as HuggingFace dataset configs and can be loaded directly with `load_dataset`.
 
 ## Corpus Scale
 
@@ -112,14 +122,39 @@ Where `{model}` uses `XxX` as a separator (e.g. `QwenXxXQwen3-Embedding-0.6B`).
 
 Loading chapter embeddings for one translation:
 
+```python
+import pyarrow.parquet as pq
+from huggingface_hub import hf_hub_download
+
+p = hf_hub_download(
+    repo_id="mrapacz/targum-corpus",
+    repo_type="dataset",
+    filename="embeddings/QwenXxXQwen3-Embedding-0.6B/language=eng/"
+             "site=ebible.org/translation=engwebp/granularity=chapter/data.parquet",
+)
+df = pq.ParquetFile(p).read().to_pandas()
+```
+
+## Similarity
+
+Pre-computed pairwise similarity scores between translations are
+distributed alongside the corpus under `similarity/`:
+
+- **`similarity/lexical/jaccard/chapter/{iso}.parquet`** -- chapter-level Jaccard token-overlap scores.
+- **`similarity/lexical/levenshtein/chapter/{iso}.parquet`** -- chapter-level Levenshtein-based scores.
+- **`similarity/semantic/cosine/{model}/chapter/{iso}.parquet`** -- chapter-level cosine similarity between embeddings (currently `QwenXxXQwen3-Embedding-8B`). A `cross_language.parquet` adds pairs that span language boundaries.
+
+One parquet file per language at chapter granularity. Total payload
+is roughly 5 GB (lexical ~3 GB, semantic/cosine ~2 GB).
+
 ```python
 import pandas as pd
 
-df = pd.read_parquet(
-    "hf://datasets/mrapacz/targum-corpus/embeddings/"
-    "QwenXxXQwen3-Embedding-0.6B/language=eng/site=ebible.org/"
-    "translation=eng-web/granularity=chapter/data.parquet"
+sem = pd.read_parquet(
+    "hf://datasets/mrapacz/targum-corpus/similarity/semantic/cosine/"
+    "QwenXxXQwen3-Embedding-8B/chapter/ita.parquet"
 )
+print(sem.head())
 ```
 
 ## Metadata
@@ -134,13 +169,15 @@ This three-layer structure lets researchers define "uniqueness" for their own ne
 
 ## Usage
 
+### Verse text
+
 ```python
 from datasets import load_dataset
 
 # Load a single translation
 ds = load_dataset(
     "mrapacz/targum-corpus",
-    data_files="corpora/ebible.org/eng/eng-web.jsonl",
+    data_files="corpora/ebible.org/eng/engwebp.jsonl",
     split="train",
 )
 print(ds[0])
@@ -153,21 +190,86 @@ ds = load_dataset(
 )
 ```
 
-Load metadata:
+### Metadata via configs
+
+Each declared config maps to one of the TSV tables and is loadable
+directly:
+
+```python
+from datasets import load_dataset
+
+works       = load_dataset("mrapacz/targum-corpus", "works",          split="train")
+editions    = load_dataset("mrapacz/targum-corpus", "editions",       split="train")
+instances   = load_dataset("mrapacz/targum-corpus", "instances",      split="train")
+book_cov    = load_dataset("mrapacz/targum-corpus", "book_coverage",  split="train")
+```
+
+Or as pandas / polars frames over the raw TSV / JSON files:
 
 ```python
 import pandas as pd
 
-editions = pd.read_csv(
-    "hf://datasets/mrapacz/targum-corpus/editions.tsv",
-    sep="\t",
+editions  = pd.read_csv("hf://datasets/mrapacz/targum-corpus/editions.tsv",  sep="\t")
+instances = pd.read_csv("hf://datasets/mrapacz/targum-corpus/instances.tsv", sep="\t")
+```
+
+### Selective access (recommended for embeddings / similarity)
+
+With ~120 GB of embeddings + ~5 GB of similarity, you almost never
+want a full clone. Pull only what you need:
+
+```python
+from huggingface_hub import hf_hub_download, snapshot_download
+
+# One specific embedding file
+path = hf_hub_download(
+    repo_id="mrapacz/targum-corpus",
+    repo_type="dataset",
+    filename="embeddings/QwenXxXQwen3-Embedding-0.6B/language=eng/"
+             "site=ebible.org/translation=engwebp/granularity=chapter/data.parquet",
+)
+
+# All English chapter embeddings for one model
+local_dir = snapshot_download(
+    repo_id="mrapacz/targum-corpus",
+    repo_type="dataset",
+    allow_patterns=[
+        "embeddings/QwenXxXQwen3-Embedding-0.6B/language=eng/**/granularity=chapter/data.parquet",
+    ],
 )
-instances = pd.read_csv(
-    "hf://datasets/mrapacz/targum-corpus/instances.tsv",
-    sep="\t",
+
+# All metadata + the public corpus, no embeddings, no similarity
+local_dir = snapshot_download(
+    repo_id="mrapacz/targum-corpus",
+    repo_type="dataset",
+    allow_patterns=["*.tsv", "*.json", "corpora/**"],
 )
 ```
 
+The repository is backed by HuggingFace's [Xet](https://huggingface.co/docs/hub/storage-backends#xet)
+storage. To pull only the slices you care about from the
+shell, use the [`hf` CLI](https://huggingface.co/docs/huggingface_hub/main/en/guides/cli)'s
+`--include` filters:
+
+```bash
+pip install -U "huggingface_hub[cli]"
+
+# One specific embedding file
+hf download mrapacz/targum-corpus --repo-type dataset \
+  --include "embeddings/QwenXxXQwen3-Embedding-0.6B/language=eng/site=ebible.org/translation=engwebp/granularity=chapter/data.parquet" \
+  --local-dir ./targum-corpus
+
+# All English chapter embeddings for one model
+hf download mrapacz/targum-corpus --repo-type dataset \
+  --include "embeddings/QwenXxXQwen3-Embedding-0.6B/language=eng/**/granularity=chapter/data.parquet" \
+  --local-dir ./targum-corpus
+
+# Metadata + public corpus only (no embeddings, no similarity)
+hf download mrapacz/targum-corpus --repo-type dataset \
+  --include "*.tsv" --include "*.json" --include "corpora/**" \
+  --local-dir ./targum-corpus
+```
+
 ## Source Data
 
 Translations were collected from 13 libraries: bible.audio, bible.com, bible.is, biblegateway.com, biblehub.com, bibles.org, biblestudytools.com, bibliepolskie.pl, crossbible.com, ebible.org, jw.org, laparola.net, obohu.cz.