add base data

.gitattributes

@@ -58,3 +58,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.csv filter=lfs diff=lfs merge=lfs -text
+*.jsonl filter=lfs diff=lfs merge=lfs -text
+*.vw filter=lfs diff=lfs merge=lfs -text
+*.native filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,136 @@
+---
+language:
+  - en
+license: cc-by-4.0
+task_categories:
+  - text-classification
+tags:
+  - topic-modeling
+  - topic-modelling
+  - text-clustering
+  - document-representation
+size_categories:
+  - 10K<n<100K
+configs:
+  - config_name: noow
+    default: true
+    data_files:
+      - split: corpus
+        path: data/noow.csv
+  - config_name: bow
+    data_files:
+      - split: corpus
+        path: data/bow.csv
+---
+
+# SciNews
+
+This TopicNet derived preparation contains the full retained population of 41,854 scientific news bodies with nine source labels, from SciNews by Dongqi Liu, Yifan Wang, Jia Loy and Vera Demberg. It is not the historical 5,000-document screening sample. Original scholarly article text and abstracts are not substituted for news bodies. This is a derived research preparation, not the original release.
+
+## Population, labels and research splits
+
+Raw duplicate identity is body text after casefold and whitespace collapse. Same-label duplicates retain the lexicographically smallest ID. Two raw duplicate groups remove two documents; eight cross-label conflict groups quarantine 16 documents. There are no identical lemma-sequence duplicate groups. Source URL and normalized DOI families are joined before splitting; one near-duplicate pair joins families without deleting its members. Original source train/test/val counts are 33,497/4,188/4,187 and are provenance, not the research memberships.
+
+Family groups combine raw/lemma duplicate relations, source URL/DOI relations where present, and body five-token shingle Jaccard similarity ≥0.9. Groups are assigned by iterative multilabel stratification with seeds 20260824 and 20260825; all retained members of a group stay together. The resulting disjoint, exhaustive research splits are **fit 25,109, calibration 8,370, validation 8,375**. The Hugging Face split `corpus` contains every retained document; it is not `fit`. Original source membership is preserved in accepted audit/documents metadata and does not define these research splits.
+
+Retained class supports come directly from `audit.json.output_rows`:
+
+| Label | Documents |
+| --- | ---: |
+| Biology | 8571 |
+| Chemistry | 3381 |
+| Computer | 1142 |
+| Earth | 3856 |
+| Medicine | 15098 |
+| Nano | 2638 |
+| Other | 1164 |
+| Physics | 4851 |
+| Space | 1153 |
+
+## Representations and pinned preparation
+
+The source adapter maps `News_Body` to raw_text, `News_Title` to separate title metadata and `Topic` to the external label. IDs are `scinews_<original-split>_<six-digit-row-index>`; their original split component does not imply research membership.
+
+Both configurations read a single UTF-8 CSV with exactly `id,raw_text,vw_text`, sorted lexicographically by ID. `raw_text` is only the saved `documents.jsonl.raw_text` body. Title text is separate metadata. NOOW VW preserves the saved prepared line exactly, excluding its line terminator: document ID followed by `|@title` followed by `|@word`. It retains every prepared occurrence, including terms outside the fitted vocabulary. BOW counts those same terms separately within each namespace as `token:count`, with lexicographic token ordering and the same namespace ordering. Class labels never appear as VW namespaces.
+
+NOOW means occurrence order after preprocessing, not untouched or reversible prose. The accepted English preprocessing pins Python 3.11.15, NLTK 3.9.2 and NLTK-data commit `550b6625bcef1f2abff2ff770a5a0d272c9c6b2a`: WordPunctTokenizer, lowercase per original span, minimum token length 2, averaged_perceptron_tagger_eng, WordNet POS mapping J/V/N/R with noun fallback, English stopword removal and `lemma.isalpha()` filtering. The copied audit records the exact stopwords and resource hashes. Names are not masked.
+
+The research model uses **only @word**, with vocabulary fitted on `fit` only, `min_df=5` and `max_df=0.9`. The saved `*-model.vw` and `*.vw` inputs project prepared text to @word; the separate fitted dictionary controls the vocabulary while the sequence retains OOV positions. The CSVs preserve full prepared sequences, including the title namespace where present. Occurrence offsets and chains are separate saved artifacts, not CSV columns.
+
+`data/accepted/documents.jsonl` stores `offsets` and `title_offsets` as `[start, end]` arrays: zero-based Python character positions with an exclusive end (not byte offsets), relative to `raw_text` and `raw_title`, respectively. The corresponding `occurrences` and `title_occurrences` records contain `lemma`, `start`, `end` and `occurrence_index`, aligned in occurrence order with `tokens`/`offsets` and `title_tokens`/`title_offsets`; a lemma need not equal the raw substring. Per-split `data/accepted/scinews/*-occurrences.jsonl` files contain TF-IDF decision records keyed by `document_id` and `occurrence_index`, rather than raw spans. Join `document_id` to the document's `id` and use `occurrence_index` to select its body occurrence and raw span.
+
+`tfidf-bank.json` and `*-chains.jsonl` preserve the immutable accepted TF-IDF reference bank and extracted chains. These are automatic reference constructions, not gold human annotations. This packaging converts saved representations and does not rerun preprocessing, dictionary creation, reference fitting, scoring or model training. Candidate model results and selected human-audit examples are not included.
+
+## Loading and metadata joins
+
+Authenticate with an account that has access if the intended remote repository is private:
+
+```python
+from datasets import load_dataset
+
+repo = "TopicNet/SciNews"
+corpus = load_dataset(repo, "noow", split="corpus", token=True)
+bow = load_dataset(repo, "bow", split="corpus", token=True)
+```
+
+Labels and research splits are auxiliary JSON files. Their actual schema is a direct ID-to-label-list object and a direct split-name-to-ID-list object, respectively; there are no `labels` or `splits` wrappers:
+
+```python
+import json
+from huggingface_hub import hf_hub_download
+
+def metadata_path(name):
+    return hf_hub_download(repo, "data/accepted/scinews/" + name,
+                           repo_type="dataset", token=True)
+
+with open(metadata_path("labels.json"), encoding="utf-8") as handle:
+    labels = json.load(handle)
+with open(metadata_path("splits.json"), encoding="utf-8") as handle:
+    splits = json.load(handle)
+membership = {doc_id: split for split, ids in splits.items() for doc_id in ids}
+annotated = corpus.map(lambda row: {"labels": labels[row["id"]],
+                                   "research_split": membership[row["id"]]})
+fit = annotated.filter(lambda row: row["research_split"] == "fit")
+```
+
+For offline access, obtain the complete package including actual LFS contents and set `HF_DATASETS_OFFLINE=1` and `HF_HUB_OFFLINE=1` before starting Python. Use a writable cache:
+
+```python
+import json
+from pathlib import Path
+from datasets import load_dataset
+
+package = Path("/path/to/SciNews")
+corpus = load_dataset(str(package), "noow", split="corpus",
+                      cache_dir="/writable/cache")
+bow = load_dataset(str(package), "bow", split="corpus",
+                   cache_dir="/writable/cache")
+with (package / "data/accepted/scinews/labels.json").open(encoding="utf-8") as handle:
+    labels = json.load(handle)
+with (package / "data/accepted/scinews/splits.json").open(encoding="utf-8") as handle:
+    splits = json.load(handle)
+membership = {doc_id: split for split, ids in splits.items() for doc_id in ids}
+fit = corpus.filter(lambda row: membership[row["id"]] == "fit")
+```
+
+An explicit CSV reader is also available:
+`load_dataset("csv", data_files={"corpus": str(package / "data/noow.csv")}, split="corpus", cache_dir="/writable/cache")`.
+For independent Python CSV inspection, raise `csv.field_size_limit` when needed for long bodies.
+
+## Accepted files and integrity
+
+The byte-preserving accepted tree totals **5,428,036,902 bytes** before adding export CSVs. NOOW CSV is **287,965,748 bytes** and BOW CSV is **259,061,878 bytes**.
+
+All 34 files from the accepted source subtree are preserved byte for byte under `data/accepted/`, including the original nested `scinews/` directory. Root accepted files include documents, audit, config, spec, diagnostics, manifest, checksum manifest and TF-IDF bank. The nested directory retains labels, splits, vocabulary, eligible labels, dictionary exports/native dictionary and per-split prepared/model VW, occurrences, summaries and chains. Saved machine paths are provenance, not loading requirements.
+
+[EXPORT_PROVENANCE.json](EXPORT_PROVENANCE.json) records each original file size/hash, input source pins, the pinned source manifest digest, verbatim source README identity, conversion rules and both CSV sizes/hashes. The original inner checksum manifests are unchanged. Root [SHA256SUMS](SHA256SUMS) covers every package file except itself and Git internals; run `sha256sum -c SHA256SUMS` from this directory. Git LFS attributes apply to CSV, JSONL, VW and native dictionaries; small JSON metadata and cards remain ordinary Git files.
+
+## Rights, limitations and citation
+
+The [official dataset card](https://huggingface.co/datasets/dongqi-me/SciNews) declares CC BY 4.0 and also imposes gated access terms restricting use to non-commercial academic research or education. Both statements are retained in the verbatim [SOURCE_README.md](SOURCE_README.md); the CC BY tag does not erase the gate conditions. Their applicability to redistribution needs clarification, and redistribution remains unresolved. News articles can carry third-party copyright and other obligations.
+
+These local packages are intended for private research storage; remote visibility has not been verified or changed. Private storage does not supply redistribution permission. Labels and topic distributions reflect source selection, historical coverage, class imbalance and automatic preprocessing. Family grouping limits known leakage but does not establish that every semantic duplicate was detected. Normalized representations do not guarantee current facts, absence of personal information or suitability for downstream decisions.
+
+Dongqi Liu, Yifan Wang, Jia Loy and Vera Demberg (2024), [*SciNews: From Scholarly Complexities to Public Narratives – A Dataset for Scientific News Report Generation*](https://aclanthology.org/2024.lrec-main.1258/), LREC-COLING. See the [project page](https://dongqi.me/projects/SciNews).
+
+Report this TopicNet derived population and the source manifest hash in [EXPORT_PROVENANCE.json](EXPORT_PROVENANCE.json). The [research repository](https://github.com/Alvant/Alekseev2026IntraTextCoherence) contains the [source adapters and preprocessing](https://github.com/Alvant/Alekseev2026IntraTextCoherence/blob/master/work/wintra/lib/dataset_adapters.py) and [model input preparation](https://github.com/Alvant/Alekseev2026IntraTextCoherence/blob/master/work/wintra/experiments/prepare_model_inputs.py). It is currently private and access-controlled; viewing those links requires permission. Code is linked rather than bundled, and this package does not promise public standalone reproducibility.

data/bow.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a37cd1f1cb47681e9a0d8a2054a56844fb1a9738eddb67668de22d06d78231d0
+size 259061878

data/noow.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40df21605e7e4dd6a8f58a49938b252329cb78781c933645d0dc5ff605078c9c
+size 287965748