Upload folder using huggingface_hub

README.md

@@ -11,18 +11,18 @@ configs:
   data_files:
   - split: train
     path: cleaned/data.parquet
-- config_name: precleaned
+- config_name: filtered
   data_files:
   - split: train
-    path: precleaned/data.parquet
+    path: filtered/data.parquet
 - config_name: scored
   data_files:
   - split: train
     path: scored/data.parquet
-- config_name: filtered
+- config_name: precleaned
   data_files:
   - split: train
-    path: filtered/data.parquet
+    path: precleaned/data.parquet
 tags:
 - bashkir
 - russian
@@ -36,114 +36,87 @@ tags:
 
 # Bashkir-Russian Wikipedia Parallel Corpus
 
-Sentence-level Bashkir-Russian parallel data extracted from the Bashkir and
-Russian Wikipedia dumps dated `2026-08-01`.
+Sentence-level Bashkir-Russian parallel dataset extracted from the corresponding Bashkir and Russian Wikipedia dumps dated `2026-08-01`.
 
-This repository provides four configurations. The default `cleaned`
-configuration is recommended for machine translation training. The
-`precleaned` configuration is an earlier, less filtered extraction provided
-for alternative preprocessing and research. The `scored` configuration adds
-LASER-based quality scores to the `cleaned` pairs. The `filtered`
-configuration applies a conservative two-encoder filter on top of `scored`.
+The dataset has been filtered with [Bashkir Language Identification (LID)](https://huggingface.co/failed09/bashkir-lid) to eliminate foreign language leakage and untranslated Russian articles. Multilingual semantic alignment is evaluated using Meta LASER3, Google LaBSE, and the specialized in-domain [Bashkir-Russian Pair Scorer](https://huggingface.co/failed09/bashkir-pair-scorer).
 
 ## Configurations
 
-### cleaned
+| Config | Pairs | Articles | Description |
+|---|---:|---:|---|
+| **`filtered`** (recommended) | **66,498** | 31,574 | High-precision parallel subset filtered by two-voice scoring (LASER3 + LaBSE), verified with LID v3 and annotated with `pair_score`. Recommended for machine translation. |
+| **`cleaned`** | **68,737** | 33,582 | Base alignment with article-level sentence mapping, Wiki markup and service boilerplate removed. Untranslated foreign leaks removed via LID v3. |
+| **`scored`** | **68,737** | 33,582 | Identical rows to `cleaned`, retaining Meta LASER margin scores and `pair_score`. |
+| **`precleaned`** | **84,512** | 40,705 | Broad extraction with relaxed sentence-length heuristics, verified with LID v3. Provided for custom filtering experiments. |
 
-- 72,007 sentence pairs
-- 34,566 linked article pairs
-- URLs, MediaWiki markup, card fragments and malformed punctuation removed
-- duplicate and identical pairs removed
-- sentence indices retained
-
-Fields:
+## Fields and Schema
 
 - `ba_title`: Bashkir Wikipedia article title
 - `ru_title`: Russian Wikipedia article title
-- `section`: `lead` or `facts`
+- `section`: Section category (`lead` or `facts`)
 - `ba`: Bashkir sentence
 - `ru`: Russian sentence
-- `ba_sentence_index`: sentence index in the Bashkir article
-- `ru_sentence_index`: sentence index in the Russian article
+- `ba_sentence_index`: Sentence index in the Bashkir article
+- `ru_sentence_index`: Sentence index in the Russian article
+- `laser_cos`: Cosine similarity between Meta LASER embeddings
+- `laser_margin`: Margin normalized by same-language kNN density
+- `laser_margin_xling`: Margin normalized by cross-lingual kNN density (Artetxe & Schwenk, 2018)
+- `labse_cos`: Cosine similarity between Google LaBSE sentence embeddings
+- `labse_margin_xling`: Cross-lingual margin score for LaBSE embeddings
+- `pair_score`: Semantic compatibility score from [Bashkir-Russian Pair Scorer](https://huggingface.co/failed09/bashkir-pair-scorer) (`medium` model, 0.0 to 1.0)
 
-### precleaned
+## Score Distributions
 
-- 91,949 sentence pairs
-- earlier extraction with lighter filtering
-- contains older Wiki markup, links and alignment noise
+Score percentiles computed over the 66,498 pairs in the `filtered` configuration:
 
-Fields:
+| Score | p01 | p05 | p25 | p50 | p75 | p95 | p99 |
+|---|---|---|---|---|---|---|---|
+| `laser_cos` | 0.666 | 0.749 | 0.833 | 0.864 | 0.888 | 0.914 | 0.928 |
+| `laser_margin_xling` | 0.845 | 0.921 | 0.994 | 1.036 | 1.065 | 1.103 | 1.130 |
+| `labse_cos` | 0.312 | 0.509 | 0.707 | 0.789 | 0.856 | 0.926 | 0.954 |
+| `labse_margin_xling` | 0.602 | 0.850 | 1.044 | 1.135 | 1.204 | 1.294 | 1.355 |
+| `pair_score` | 0.039 | 0.394 | 0.950 | 0.987 | 0.996 | 0.999 | 0.999 |
 
-- `ba_title`, `ru_title`, `section`, `ba`, `ru`
+## Source and Processing
 
-### scored
+The data was extracted from corresponding Bashkir and Russian Wikipedia articles using the `bawiki-20260801` and `ruwiki-20260801` dumps.
 
-- 72,007 sentence pairs — **identical rows to `cleaned`**, nothing added or removed
-- quality scores computed with Meta LASER sentence embeddings:
-  - `laser_cos`: cosine similarity between LASER embeddings of the pair
-  - `laser_margin`: margin normalized by same-language kNN density
-  - `laser_margin_xling`: margin normalized by **cross-lingual** kNN density
-    (Artetxe & Schwenk, 2018) — the recommended LASER filter score
+1. **Language Gate:** Every candidate sentence pair is verified with [Bashkir Language Identification (LID)](https://huggingface.co/failed09/bashkir-lid) (`ba_prob >= 0.50`) to remove untranslated Russian articles and foreign language fragments that entered Wikipedia via automated templates.
+2. **Text Sanitization:** MediaWiki markup, table formatting, citation markers and broken punctuation are stripped.
+3. **Multilingual Scoring:** Parallel candidates are scored with LASER3 (`bak_Cyrl.v1` for Bashkir, `laser2` for Russian), Google LaBSE, and [Bashkir-Russian Pair Scorer](https://huggingface.co/failed09/bashkir-pair-scorer) (`medium` model).
 
-Encoders: LASER3 `bak_Cyrl` (NLLB) for Bashkir, LASER2 for Russian,
-1024-dim L2-normalized embeddings, k=4 neighbours
-(`laser_encoders`, facebookresearch/LASER).
+## Loading
 
-### filtered
+```python
+from datasets import load_dataset
 
-- **69,706 sentence pairs** (−3.20% from `scored`)
-- conservative two-encoder filter: a pair is removed only when **both**
-  scorers place it in their bottom 5%:
-  `laser_margin_xling < 0.876 AND labse_margin_xling < 0.723`
-- second encoder: **LaBSE** (Google, 481M parameters, fp16) — independent
-  architecture and training data, so its errors are largely uncorrelated
-  with LASER's
-- all score columns retained: `laser_cos`, `laser_margin`,
-  `laser_margin_xling`, `labse_cos`, `labse_margin_xling`
-- removed pairs are dominated by template mismatch sentences (river-tributary
-  stubs with different facts) and cross-article topic matches
+# Recommended subset for machine translation training
+filtered = load_dataset("failed09/bashkir-wikipedia-parallel", "filtered")
 
-Score distributions over the full 72,007 pairs:
+# Custom filtering example combining pair_score and LASER margin
+df = filtered["train"].to_pandas()
+strict_subset = df[(df.pair_score >= 0.80) & (df.laser_margin_xling >= 0.95)]
+print(f"Strict high-quality subset: {len(strict_subset):,} sentence pairs")
+```
 
-| score | p01 | p05 | p25 | p50 | p75 | p95 | p99 |
-|---|---|---|---|---|---|---|---|
-| `laser_cos` | 0.597 | 0.705 | 0.828 | 0.863 | 0.887 | 0.914 | 0.928 |
-| `laser_margin_xling` | 0.760 | 0.876 | 0.988 | 1.033 | 1.063 | 1.102 | 1.129 |
-| `labse_cos` | 0.206 | 0.408 | 0.698 | 0.789 | 0.859 | 0.932 | 0.963 |
-| `labse_margin_xling` | 0.373 | 0.723 | 1.035 | 1.132 | 1.202 | 1.295 | 1.358 |
+## Related Resources
 
-## Source and Processing
+- [Bashkir Wikipedia Monolingual Corpus](https://huggingface.co/datasets/failed09/bashkir-wikipedia-monolingual): Monolingual companion corpus with 800K cleaned sentences extracted from the same Wikipedia dump.
+- [Bashkir-Russian Pair Scorer](https://huggingface.co/failed09/bashkir-pair-scorer): In-domain cross-encoder models (`nano`, `mini`, `medium`) used to generate `pair_score`.
+- [Bashkir Language Identification (LID)](https://huggingface.co/failed09/bashkir-lid): The standalone ONNX LID model used to filter non-Bashkir text.
+- [Bashkir Frequency Index](https://huggingface.co/datasets/failed09/bashkir-frequency-index): Canonical word-frequency and coverage statistics.
+- [Bashkir Word N-gram Index](https://huggingface.co/datasets/failed09/bashkir-ngram-index): Unigram, bigram and trigram sequence counts.
 
-The data was extracted from corresponding Bashkir and Russian Wikipedia
-articles using the `bawiki-20260801` and `ruwiki-20260801` dumps.
+## Quality and Use
 
-The cleaned configuration uses article-local sentence indices, numeric
-anchors for factual pairs, length-ratio checks, language checks, duplicate
-removal and service-fragment filtering. The `scored` and `filtered`
-configurations add multilingual sentence-embedding scoring on top of the
-same alignment without modifying any text.
+This dataset is compiled from crowd-sourced Wikipedia articles. While automated language identification, regex sanitization, and dual-encoder filtering significantly reduce noise, users should expect occasional translation mismatches, differing article lengths, or encyclopedic jargon. Always inspect and validate the alignment before production deployment.
 
 ## License
 
-The source text is derived from Wikipedia and is distributed under the
-[Creative Commons Attribution-ShareAlike 4.0 International license](https://creativecommons.org/licenses/by-sa/4.0/).
-
-Please preserve Wikipedia attribution when redistributing or publishing
-derivative datasets.
+The underlying source text is derived from Wikipedia and is distributed under the
+[Creative Commons Attribution-ShareAlike 4.0 International license (CC BY-SA 4.0)](https://creativecommons.org/licenses/by-sa/4.0/).
 
-## Loading
-
-```python
-from datasets import load_dataset
-
-cleaned = load_dataset("failed09/bashkir-wikipedia-parallel", "cleaned")
-scored = load_dataset("failed09/bashkir-wikipedia-parallel", "scored")
-filtered = load_dataset("failed09/bashkir-wikipedia-parallel", "filtered")
-
-# strict MT-training subset with your own threshold
-df = scored["train"].to_pandas()
-strict = df[(df.laser_margin_xling >= 0.90) & (df.labse_margin_xling >= 1.0)]
-```
+Please preserve Wikipedia attribution when redistributing or publishing derivative models and datasets.
 
 ## Citation
 
@@ -153,6 +126,13 @@ strict = df[(df.laser_margin_xling >= 0.90) & (df.labse_margin_xling >= 1.0)]
   author = {failed09},
   year = {2026},
   publisher = {Hugging Face},
-  url = {https://huggingface.co/datasets/failed09/bashkir-wikipedia-parallel}
+  url = {https://huggingface.co/datasets/failed09/bashkir-wikipedia-parallel},
+  note = {Cleaned Bashkir-Russian Wikipedia parallel corpus with LASER, LaBSE and Pair Scorer metrics}
 }
 ```
+
+## Open Bashkir Data and Sources 🐝
+
+This dataset is part of an open-source initiative to advance NLP resources, machine translation, and language technologies for the Bashkir language. Related models, benchmarks and datasets are available on the author's Hugging Face profile.
+
+The author does not claim copyright ownership of the original Wikipedia texts. All rights and licensing remain with the respective Wikipedia contributors and the Wikimedia Foundation under CC BY-SA 4.0.

SHA256SUMS

@@ -0,0 +1,6 @@
+a1efe518e6ede30463089d83808052e6b2e84c0323ed97f0a6dbe39a8b504341  README.md
+00d32007eddd9dfcef26027e18f33caeb8f6866371a8c96f9b1060c6be8a266f  metadata.json
+dbece47e97fa44fccba19d31ddd574cd069e0d843755c889ed757fe4e0dd0e62  cleaned/data.parquet
+5473b45a4a0c3949c4539d2b3f4bbd6baaa676fa4ae2149e5878fefb5667fdca  filtered/data.parquet
+842b68ff104f7befd20dbde56bc471d3e315100fc5cba1ab582dac16e6bdfff4  precleaned/data.parquet
+e479b053d5b05f1aa4745ca9602d95b97f4fc7ba7f551717320b20da41396340  scored/data.parquet

cleaned/data.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6c336138bc85d1c88066de1707de7e63331693c584ffdee0e980785bca855753
-size 16422979
+oid sha256:dbece47e97fa44fccba19d31ddd574cd069e0d843755c889ed757fe4e0dd0e62
+size 15834185

filtered/data.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dce11dc537c894976e13986ebb89f52337999a29e5db946a3abc19b2ddd76988
-size 18793907
+oid sha256:5473b45a4a0c3949c4539d2b3f4bbd6baaa676fa4ae2149e5878fefb5667fdca
+size 18220398

metadata.json

@@ -6,42 +6,41 @@
   "configs": {
     "cleaned": {
       "file": "cleaned/data.parquet",
-      "version": "v6",
-      "rows": 72007,
-      "linked_article_pairs": 34566,
-      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index"]
+      "version": "v8.1",
+      "rows": 68737,
+      "linked_article_pairs": 33582,
+      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index"],
+      "filter": "LID v3 verified (ba_prob >= 0.50); foreign text, duplicate and empty pairs removed"
     },
     "precleaned": {
       "file": "precleaned/data.parquet",
-      "version": "v1",
-      "rows": 91949,
-      "columns": ["ba_title", "ru_title", "section", "ba", "ru"]
+      "version": "v1.1",
+      "rows": 84512,
+      "linked_article_pairs": 40705,
+      "columns": ["ba_title", "ru_title", "section", "ba", "ru"],
+      "filter": "LID v3 verified (ba_prob >= 0.50)"
     },
     "scored": {
       "file": "scored/data.parquet",
-      "version": "v7",
-      "rows": 72007,
-      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index", "laser_cos", "laser_margin", "laser_margin_xling"],
+      "version": "v8.1",
+      "rows": 68737,
+      "linked_article_pairs": 33582,
+      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index", "laser_cos", "laser_margin", "laser_margin_xling", "pair_score"],
       "scores": {
-        "method": "LASER margin-based bitext scoring (Artetxe & Schwenk 2018)",
-        "encoder_ba": "laser3-bak_Cyrl.v1",
-        "encoder_ru": "laser2",
-        "embedding_dim": 1024,
-        "knn_k": 4,
-        "note": "laser_margin uses same-language kNN, laser_margin_xling uses cross-lingual kNN and is the recommended LASER filter score"
+        "laser": "LASER margin-based bitext scoring (Artetxe & Schwenk 2018), laser3-bak_Cyrl.v1 + laser2",
+        "pair_score": "Bashkir-Russian Pair Scorer medium (ONNX fast128 cross-encoder compatibility score)"
       }
     },
     "filtered": {
       "file": "filtered/data.parquet",
-      "version": "v8",
-      "rows": 69706,
-      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index", "laser_cos", "laser_margin", "laser_margin_xling", "labse_cos", "labse_margin_xling"],
+      "version": "v8.1",
+      "rows": 66498,
+      "linked_article_pairs": 31574,
+      "columns": ["ba_title", "ru_title", "section", "ba", "ru", "ba_sentence_index", "ru_sentence_index", "laser_cos", "laser_margin", "laser_margin_xling", "labse_cos", "labse_margin_xling", "pair_score"],
       "filter": {
-        "name": "two_voice_both_low",
-        "rule": "drop if laser_margin_xling < 0.876 AND labse_margin_xling < 0.723 (bottom 5% of both scorers)",
-        "rows_dropped": 2301,
-        "encoders": ["laser3-bak_Cyrl.v1 + laser2 (LASER3/LASER2, dim 1024)", "sentence-transformers/LaBSE (dim 768, fp16)"],
-        "design": "pairs removed only when two independent encoders agree on rejection; all score columns retained for auditability"
+        "name": "lid_plus_two_voice",
+        "rule": "LID v3 (ba_prob >= 0.50) + two-voice (drop if laser_margin_xling < 0.876 AND labse_margin_xling < 0.723)",
+        "encoders": ["laser3-bak_Cyrl.v1 + laser2 (dim 1024)", "sentence-transformers/LaBSE (dim 768)", "bashkir-pair-scorer medium"]
       }
     }
   }

precleaned/data.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:61e960e1a122d45a5ae92b6d8116590cfb0b800bb54a815b2e3721406b2fa7a6
-size 20028563
+oid sha256:842b68ff104f7befd20dbde56bc471d3e315100fc5cba1ab582dac16e6bdfff4
+size 18789138

scored/data.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:021149d2899d06512ea0e20cb203a5ebf190e856ac96f1f18854c8c9f0e246c4
-size 18058017
+oid sha256:e479b053d5b05f1aa4745ca9602d95b97f4fc7ba7f551717320b20da41396340
+size 17533851