Vasily Alexeev commited on
Commit
71433f1
·
1 Parent(s): 6927795

Selected -> Thematized

Browse files
Files changed (1) hide show
  1. README.md +14 -14
README.md CHANGED
@@ -15,7 +15,7 @@ configs:
15
  path: data/topics.jsonl.gz
16
  ---
17
 
18
- # Russian Wikipedia quality articles with subjects — 20261001
19
 
20
  This Russian-language collection contains **17,251 Wikipedia articles**
21
  with at least one subject label from a fixed taxonomy of **28 subjects**.
@@ -83,7 +83,7 @@ Of the **17,413 source articles**, **14,945** required no additional review:
83
  **2,468** articles had evidence from **44 ambiguous editorial rubrics** and
84
  received additional AI-assisted review. Of these, **2,306** were retained and
85
  **162** were excluded because their final subject lists were empty. In total,
86
- **2,054** articles had their subject lists changed. The final Selected collection
87
  contains **17,251 articles: 15,599 with one subject and 1,652 with several**.
88
  Additional review records a step in annotation: it does not imply an automatic
89
  label change and is not a confidence score. Articles without additional review
@@ -181,7 +181,7 @@ The repository does not store a second permanent copy as VW text files.
181
  The NLP preparation section describes the normalized namespaces and text fields.
182
 
183
  The sidecar [data/labels.json](data/labels.json) is a JSON object mapping every
184
- Selected article ID to its list of subject slugs, matching both article `topics`
185
  and the CSV `@topic` namespace. It is not a separate dataset configuration.
186
 
187
  ```python
@@ -195,7 +195,7 @@ labels = json.loads(Path("data/labels.json").read_text(encoding="utf-8"))
195
 
196
  [data/topic-provenance.jsonl.gz](data/topic-provenance.jsonl.gz) contains one
197
  record for each of the **17,413 Full article IDs**, including the 162 articles
198
- excluded from Selected. Records are ordered by numeric page ID. This is frozen
199
  annotation metadata, separate from corpus reconstruction and NLP build outputs;
200
  preprocessing does not re-annotate it or create new subject evidence. It records
201
  the initial specific-rubric assignments and the final accepted assignments,
@@ -205,11 +205,11 @@ without publishing internal annotation decision logs.
205
  | --- | --- | --- |
206
  | `id` | string | Full article ID, `ruwiki/<page_id>`; join to Full to inspect excluded articles. |
207
  | `initial_topics` | array of strings | Subjects supported by specific direct editorial rubrics before additional review. |
208
- | `topics` | array of strings | Final subjects; matches `data/labels.json` for every Selected ID. An empty list identifies an excluded Full article. |
209
  | `additional_review` | boolean | Whether baseline ambiguous-rubric evidence led to additional AI-assisted review; not whether labels changed or a confidence score. |
210
  | `unresolved_rubrics` | array of strings | Ambiguous direct editorial rubrics at the initial stage; retained as historical evidence even after review, not a claim that the final subjects remain unresolved. |
211
 
212
- To stream only reviewed articles retained in Selected, run this from the root of
213
  the downloaded repository. Use `not row["topics"]` instead of `row["topics"]`
214
  to inspect the excluded articles, joining their IDs to Full:
215
 
@@ -358,7 +358,7 @@ can inspect these losses or apply a different extraction procedure. Neither
358
  The [standalone preparation code](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026/tree/main/preprocessing)
359
  and [full instructions](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026#nlp-preparation)
360
  are maintained in Full. Run the preparation commands from Full's
361
- `preprocessing/` directory, including when preparing only Selected. Local
362
  repository directories may have any names. Replace the paths below with your own
363
  absolute paths: `work_dir` is a separate working directory, and `selected_data`
364
  points to this collection's `data/` directory. Choose fresh output and scratch
@@ -400,7 +400,7 @@ for a new build. Full's instructions explain reuse of local JSONL files and
400
  access to restricted source repositories.
401
 
402
  After the structural NLP checks, the shared entry point independently compares
403
- CSV and Selected label byte sizes and SHA-256 hashes with published expectations
404
  in the corpus contract, reporting `PUBLISHED_NLP_OK`. It does not rely on the
405
  new build manifest as its authority and does not replay NLP normalization.
406
 
@@ -429,7 +429,7 @@ progress; the manifest reports empty `@word` articles by count and ordered IDs
429
  without dropping them.
430
 
431
  Bigrams and trigrams use independent NLTK PMI selection after frequency ≥5,
432
- with `nbest(32100)[100:]` for each order. This selection uses **all Selected
433
  article bodies**, independently of Full; it is corpus-global rather than fitted
434
  only on a future training partition. For text-only experiments, use only `@word`
435
  to avoid exposing category, title or subject metadata to the model. The workflow
@@ -451,12 +451,12 @@ bow = load_dataset(
451
  )
452
  ```
453
 
454
- A measured NLP build of **both Full and Selected** from ready JSONL snapshots took
455
  42 min 5.83 s, with 10.58 GiB peak resident memory, approximately 3.74 GiB of final
456
  output and 0.53 GiB of scratch. These are measurements from one environment,
457
- not Selected-only estimates or measurements of replay from raw Wikimedia dumps.
458
 
459
- ## Rebuild Selected from source data
460
 
461
  Full's [corpus replay instructions](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026#rebuild-the-corpus-from-wikimedia-dumps)
462
  cover the six pinned 20261001 dumps, explicit download resumption with `--resume`,
@@ -489,7 +489,7 @@ annotation directory must already contain this collection's `labels.json` and
489
  `topics.jsonl.gz`. Reconstruction retains the published mapping and does not
490
  rerun AI annotation or establish new evidence for label correctness.
491
 
492
- If a previously accepted Full corpus is already available, build only Selected
493
  without rereading the dumps. From the same Full `preprocessing/` directory,
494
  resolve Full's data path:
495
 
@@ -506,7 +506,7 @@ python prepare.py --stage corpus --action verify --dataset selected \
506
  --output-dir "$work_dir/selected-output"
507
  ```
508
 
509
- This Selected-only workflow has been verified byte for byte against the published
510
  archive of all 17,251 articles, the labels file and the topic definitions.
511
 
512
  Use fresh output and scratch directories. Corpus verification checks exact
 
15
  path: data/topics.jsonl.gz
16
  ---
17
 
18
+ # RuWiki-Good-Thematized-2026 — Russian Wikipedia quality articles with subjects
19
 
20
  This Russian-language collection contains **17,251 Wikipedia articles**
21
  with at least one subject label from a fixed taxonomy of **28 subjects**.
 
83
  **2,468** articles had evidence from **44 ambiguous editorial rubrics** and
84
  received additional AI-assisted review. Of these, **2,306** were retained and
85
  **162** were excluded because their final subject lists were empty. In total,
86
+ **2,054** articles had their subject lists changed. The final Thematized collection
87
  contains **17,251 articles: 15,599 with one subject and 1,652 with several**.
88
  Additional review records a step in annotation: it does not imply an automatic
89
  label change and is not a confidence score. Articles without additional review
 
181
  The NLP preparation section describes the normalized namespaces and text fields.
182
 
183
  The sidecar [data/labels.json](data/labels.json) is a JSON object mapping every
184
+ Thematized article ID to its list of subject slugs, matching both article `topics`
185
  and the CSV `@topic` namespace. It is not a separate dataset configuration.
186
 
187
  ```python
 
195
 
196
  [data/topic-provenance.jsonl.gz](data/topic-provenance.jsonl.gz) contains one
197
  record for each of the **17,413 Full article IDs**, including the 162 articles
198
+ excluded from Thematized. Records are ordered by numeric page ID. This is frozen
199
  annotation metadata, separate from corpus reconstruction and NLP build outputs;
200
  preprocessing does not re-annotate it or create new subject evidence. It records
201
  the initial specific-rubric assignments and the final accepted assignments,
 
205
  | --- | --- | --- |
206
  | `id` | string | Full article ID, `ruwiki/<page_id>`; join to Full to inspect excluded articles. |
207
  | `initial_topics` | array of strings | Subjects supported by specific direct editorial rubrics before additional review. |
208
+ | `topics` | array of strings | Final subjects; matches `data/labels.json` for every Thematized ID. An empty list identifies an excluded Full article. |
209
  | `additional_review` | boolean | Whether baseline ambiguous-rubric evidence led to additional AI-assisted review; not whether labels changed or a confidence score. |
210
  | `unresolved_rubrics` | array of strings | Ambiguous direct editorial rubrics at the initial stage; retained as historical evidence even after review, not a claim that the final subjects remain unresolved. |
211
 
212
+ To stream only reviewed articles retained in Thematized, run this from the root of
213
  the downloaded repository. Use `not row["topics"]` instead of `row["topics"]`
214
  to inspect the excluded articles, joining their IDs to Full:
215
 
 
358
  The [standalone preparation code](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026/tree/main/preprocessing)
359
  and [full instructions](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026#nlp-preparation)
360
  are maintained in Full. Run the preparation commands from Full's
361
+ `preprocessing/` directory, including when preparing only Thematized. Local
362
  repository directories may have any names. Replace the paths below with your own
363
  absolute paths: `work_dir` is a separate working directory, and `selected_data`
364
  points to this collection's `data/` directory. Choose fresh output and scratch
 
400
  access to restricted source repositories.
401
 
402
  After the structural NLP checks, the shared entry point independently compares
403
+ CSV and Thematized label byte sizes and SHA-256 hashes with published expectations
404
  in the corpus contract, reporting `PUBLISHED_NLP_OK`. It does not rely on the
405
  new build manifest as its authority and does not replay NLP normalization.
406
 
 
429
  without dropping them.
430
 
431
  Bigrams and trigrams use independent NLTK PMI selection after frequency ≥5,
432
+ with `nbest(32100)[100:]` for each order. This selection uses **all Thematized
433
  article bodies**, independently of Full; it is corpus-global rather than fitted
434
  only on a future training partition. For text-only experiments, use only `@word`
435
  to avoid exposing category, title or subject metadata to the model. The workflow
 
451
  )
452
  ```
453
 
454
+ A measured NLP build of **both Full and Thematized** from ready JSONL snapshots took
455
  42 min 5.83 s, with 10.58 GiB peak resident memory, approximately 3.74 GiB of final
456
  output and 0.53 GiB of scratch. These are measurements from one environment,
457
+ not Thematized-only estimates or measurements of replay from raw Wikimedia dumps.
458
 
459
+ ## Rebuild Thematized from source data
460
 
461
  Full's [corpus replay instructions](https://huggingface.co/datasets/TopicNet/RuWiki-Good-2026#rebuild-the-corpus-from-wikimedia-dumps)
462
  cover the six pinned 20261001 dumps, explicit download resumption with `--resume`,
 
489
  `topics.jsonl.gz`. Reconstruction retains the published mapping and does not
490
  rerun AI annotation or establish new evidence for label correctness.
491
 
492
+ If a previously accepted Full corpus is already available, build only Thematized
493
  without rereading the dumps. From the same Full `preprocessing/` directory,
494
  resolve Full's data path:
495
 
 
506
  --output-dir "$work_dir/selected-output"
507
  ```
508
 
509
+ This Thematized-only workflow has been verified byte for byte against the published
510
  archive of all 17,251 articles, the labels file and the topic definitions.
511
 
512
  Use fresh output and scratch directories. Corpus verification checks exact