Add filtered config (two-voice LASER+LaBSE filter, v8): README.md

README.md

@@ -19,6 +19,10 @@ configs:
   data_files:
   - split: train
     path: scored/data.parquet
+- config_name: filtered
+  data_files:
+  - split: train
+    path: filtered/data.parquet
 tags:
 - bashkir
 - russian
@@ -35,12 +39,12 @@ tags:
 Sentence-level Bashkir-Russian parallel data extracted from the Bashkir and
 Russian Wikipedia dumps dated `2026-08-01`.
 
-This repository provides three configurations. The default `cleaned`
+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 so that users can apply
-their own filtering thresholds.
+LASER-based quality scores to the `cleaned` pairs. The `filtered`
+configuration applies a conservative two-encoder filter on top of `scored`.
 
 ## Configurations
 
@@ -70,60 +74,43 @@ Fields:
 
 Fields:
 
-- `ba_title`
-- `ru_title`
-- `section`
-- `ba`
-- `ru`
+- `ba_title`, `ru_title`, `section`, `ba`, `ru`
 
 ### scored
 
 - 72,007 sentence pairs — **identical rows to `cleaned`**, nothing added or removed
-- three added quality-score columns computed with Meta LASER sentence embeddings
-- lets you apply your own filtering threshold instead of trusting a fixed cut
-
-Added fields:
-
-- `laser_cos`: cosine similarity between the LASER embeddings of the Bashkir
-  and Russian sentences (higher = closer translation)
-- `laser_margin`: margin score normalized by same-language kNN density
-  (kept for completeness)
-- `laser_margin_xling`: margin score normalized by **cross-lingual** kNN
-  density in the Artetxe & Schwenk (2018) sense — the recommended filter score
-
-## How the scores were computed
-
-- Bashkir encoder: **LASER3 `bak_Cyrl`** (NLLB distillation), Russian encoder:
-  **LASER2**; both produce 1024-dim L2-normalized sentence embeddings
-  (`laser_encoders`, facebookresearch/LASER).
-- For every pair `(x, y)` with cosine similarity `cos(x, y)`:
-  `margin = 2 * cos(x, y) / (mean_{t in kNN(x)} cos(x, t) + mean_{t in kNN(y)} cos(y, t))`
-  where `kNN(x)` are the k=4 nearest neighbours of `x` among **Russian**
-  sentences and `kNN(y)` the k=4 nearest neighbours of `y` among **Bashkir**
-  sentences.
-- The margin penalizes sentences that sit in dense near-duplicate clusters.
-  Such pairs can be perfectly correct yet receive a low score — do not use a
-  high threshold blindly.
-
-## Score distributions (72,007 pairs)
+- 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
+
+Encoders: LASER3 `bak_Cyrl` (NLLB) for Bashkir, LASER2 for Russian,
+1024-dim L2-normalized embeddings, k=4 neighbours
+(`laser_encoders`, facebookresearch/LASER).
+
+### filtered
+
+- **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
+
+Score distributions over the full 72,007 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 |
-
-## Recommended thresholds
-
-- **Strict (MT training):** drop pairs with
-  `laser_margin_xling < 0.90 OR laser_cos < 0.70` → removes 6.6% (4,778 pairs),
-  keeps 67,229.
-- **Conservative:** drop pairs with
-  `laser_margin_xling < 0.85 OR laser_cos < 0.65` → removes 3.8% (2,750 pairs),
-  keeps 69,257.
-- **Research:** keep all rows and analyze the score distribution directly.
-
-The alignment is automatic and has not been fully verified by human
-annotators; the scores make the residual noise auditable rather than hiding it.
+| `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 |
 
 ## Source and Processing
 
@@ -132,9 +119,9 @@ articles using the `bawiki-20260801` and `ruwiki-20260801` dumps.
 
 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` configuration adds
-multilingual sentence-embedding scores on top of the same alignment without
-modifying any row.
+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.
 
 ## License
 
@@ -150,13 +137,12 @@ derivative datasets.
 from datasets import load_dataset
 
 cleaned = load_dataset("failed09/bashkir-wikipedia-parallel", "cleaned")
-precleaned = load_dataset("failed09/bashkir-wikipedia-parallel", "precleaned")
 scored = load_dataset("failed09/bashkir-wikipedia-parallel", "scored")
+filtered = load_dataset("failed09/bashkir-wikipedia-parallel", "filtered")
 
-# strict MT-training filter
+# strict MT-training subset with your own threshold
 df = scored["train"].to_pandas()
-keep = (df.laser_margin_xling >= 0.90) & (df.laser_cos >= 0.70)
-mt_train = df[keep]
+strict = df[(df.laser_margin_xling >= 0.90) & (df.labse_margin_xling >= 1.0)]
 ```
 
 ## Citation