Upload README.md with huggingface_hub

README.md

@@ -10,10 +10,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 332627048
     num_examples: 87536
-  download_size: 152789447
-  dataset_size: 332627048
 - config_name: go-v1-pair-2M
   features:
   - name: query
@@ -22,10 +19,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 855194988
     num_examples: 1532319
-  download_size: 371361455
-  dataset_size: 855194988
 - config_name: java-v1-hard-negatives-100k
   features:
   - name: query
@@ -36,10 +30,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 327677194
     num_examples: 80906
-  download_size: 121324000
-  dataset_size: 327677194
 - config_name: java-v1-pair-2M
   features:
   - name: query
@@ -48,10 +39,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 1015086964
     num_examples: 1456651
-  download_size: 396953392
-  dataset_size: 1015086964
 - config_name: javascript-v1-hard-negatives-100k
   features:
   - name: query
@@ -62,10 +50,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 408914607
     num_examples: 79484
-  download_size: 176447985
-  dataset_size: 408914607
 - config_name: javascript-v1-pair-2M
   features:
   - name: query
@@ -74,10 +59,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 1084716629
     num_examples: 1298745
-  download_size: 467742815
-  dataset_size: 1084716629
 - config_name: php-v1-hard-negatives-100k
   features:
   - name: query
@@ -88,10 +70,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 368842521
     num_examples: 74945
-  download_size: 149769432
-  dataset_size: 368842521
 - config_name: php-v1-pair-2M
   features:
   - name: query
@@ -100,10 +79,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 993459020
     num_examples: 1311302
-  download_size: 405251406
-  dataset_size: 993459020
 - config_name: python-v1-hard-negatives-100k
   features:
   - name: query
@@ -114,10 +90,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 417682840
     num_examples: 96670
-  download_size: 181218116
-  dataset_size: 417682840
 - config_name: python-v1-pair-2M
   features:
   - name: query
@@ -126,10 +99,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 1338680922
     num_examples: 1779891
-  download_size: 593332698
-  dataset_size: 1338680922
 - config_name: ruby-v1-hard-negatives-100k
   features:
   - name: query
@@ -140,10 +110,7 @@ dataset_info:
     list: string
   splits:
   - name: train
-    num_bytes: 162232900
     num_examples: 67937
-  download_size: 75638556
-  dataset_size: 162232900
 - config_name: ruby-v1-pair-2M
   features:
   - name: query
@@ -152,10 +119,7 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 541359997
     num_examples: 1145414
-  download_size: 245183601
-  dataset_size: 541359997
 configs:
 - config_name: go-v1-hard-negatives-100k
   data_files:
@@ -224,7 +188,8 @@ If you need large-scale data (not samples), please refer to the original CoRNSta
 
 This release keeps the original subset layout (6 languages x pair + hard-negatives) and applies deterministic rule-based filtering.
 
-In this revision, query-level deduplication is also applied per subset: if `query` is duplicated, only the first row is kept.
+In this revision, query-level deduplication and positive-document (`pos`) deduplication are applied per subset.
+For both rules, if duplicates exist, only the first row is kept.
 
 ## Config Layout And Schema
 
@@ -236,50 +201,55 @@ Schema:
 - Pair configs: `query`, `pos`
 - Hard-negative configs: `query`, `pos`, `negs` (list[string])
 
-## Subsets And Row Counts (After Filter + Query Dedup)
+## Subsets And Row Counts (After Filter + Query Dedup + Pos Dedup)
 
 Note: subset names keep the original `2M` / `100k` naming convention for compatibility, but
 actual row counts vary by language and can be substantially lower after quality filtering and
-query-level deduplication.
+deduplication.
 
 | Subset (config name) | split | num_examples |
 | --- | --- | ---: |
-| `go-v1-pair-2M` | `train` | 1,541,111 |
-| `go-v1-hard-negatives-100k` | `train` | 87,647 |
-| `java-v1-pair-2M` | `train` | 1,491,655 |
-| `java-v1-hard-negatives-100k` | `train` | 81,657 |
-| `javascript-v1-pair-2M` | `train` | 1,310,965 |
-| `javascript-v1-hard-negatives-100k` | `train` | 79,684 |
-| `php-v1-pair-2M` | `train` | 1,343,442 |
-| `php-v1-hard-negatives-100k` | `train` | 75,632 |
-| `python-v1-pair-2M` | `train` | 1,807,480 |
-| `python-v1-hard-negatives-100k` | `train` | 97,147 |
-| `ruby-v1-pair-2M` | `train` | 1,175,219 |
-| `ruby-v1-hard-negatives-100k` | `train` | 68,382 |
+| `go-v1-pair-2M` | `train` | 1,532,319 |
+| `go-v1-hard-negatives-100k` | `train` | 87,536 |
+| `java-v1-pair-2M` | `train` | 1,456,651 |
+| `java-v1-hard-negatives-100k` | `train` | 80,906 |
+| `javascript-v1-pair-2M` | `train` | 1,298,745 |
+| `javascript-v1-hard-negatives-100k` | `train` | 79,484 |
+| `php-v1-pair-2M` | `train` | 1,311,302 |
+| `php-v1-hard-negatives-100k` | `train` | 74,945 |
+| `python-v1-pair-2M` | `train` | 1,779,891 |
+| `python-v1-hard-negatives-100k` | `train` | 96,670 |
+| `ruby-v1-pair-2M` | `train` | 1,145,414 |
+| `ruby-v1-hard-negatives-100k` | `train` | 67,937 |
 
 Total rows:
-- Pair: 8,669,872
-- Hard-negatives: 490,149
-- Overall: 9,160,021
+- Pair: 8,524,322
+- Hard-negatives: 487,478
+- Overall: 9,011,800
 
-## Filter Impact (Query Dedup Stage)
+## Dedup Impact (Pos Dedup Stage)
 
-The table below shows only the query-dedup impact on top of the previous rule-based filter.
+The table below shows the `pos`-dedup impact on top of the previous filtered + query-deduped data.
 
 | Subset | before | after | removed | removed_ratio |
 | --- | ---: | ---: | ---: | ---: |
-| `go-v1-pair-2M` | 1,992,985 | 1,541,111 | 451,874 | 22.67% |
-| `go-v1-hard-negatives-100k` | 99,663 | 87,647 | 12,016 | 12.06% |
-| `java-v1-pair-2M` | 1,752,593 | 1,491,655 | 260,938 | 14.89% |
-| `java-v1-hard-negatives-100k` | 87,504 | 81,657 | 5,847 | 6.68% |
-| `javascript-v1-pair-2M` | 1,960,276 | 1,310,965 | 649,311 | 33.12% |
-| `javascript-v1-hard-negatives-100k` | 98,025 | 79,684 | 18,341 | 18.71% |
-| `php-v1-pair-2M` | 1,710,537 | 1,343,442 | 367,095 | 21.46% |
-| `php-v1-hard-negatives-100k` | 85,460 | 75,632 | 9,828 | 11.50% |
-| `python-v1-pair-2M` | 1,990,051 | 1,807,480 | 182,571 | 9.17% |
-| `python-v1-hard-negatives-100k` | 99,535 | 97,147 | 2,388 | 2.40% |
-| `ruby-v1-pair-2M` | 1,583,047 | 1,175,219 | 407,828 | 25.76% |
-| `ruby-v1-hard-negatives-100k` | 79,040 | 68,382 | 10,658 | 13.48% |
+| `go-v1-pair-2M` | 1,541,111 | 1,532,319 | 8,792 | 0.57% |
+| `go-v1-hard-negatives-100k` | 87,647 | 87,536 | 111 | 0.13% |
+| `java-v1-pair-2M` | 1,491,655 | 1,456,651 | 35,004 | 2.35% |
+| `java-v1-hard-negatives-100k` | 81,657 | 80,906 | 751 | 0.92% |
+| `javascript-v1-pair-2M` | 1,310,965 | 1,298,745 | 12,220 | 0.93% |
+| `javascript-v1-hard-negatives-100k` | 79,684 | 79,484 | 200 | 0.25% |
+| `php-v1-pair-2M` | 1,343,442 | 1,311,302 | 32,140 | 2.39% |
+| `php-v1-hard-negatives-100k` | 75,632 | 74,945 | 687 | 0.91% |
+| `python-v1-pair-2M` | 1,807,480 | 1,779,891 | 27,589 | 1.53% |
+| `python-v1-hard-negatives-100k` | 97,147 | 96,670 | 477 | 0.49% |
+| `ruby-v1-pair-2M` | 1,175,219 | 1,145,414 | 29,805 | 2.54% |
+| `ruby-v1-hard-negatives-100k` | 68,382 | 67,937 | 445 | 0.65% |
+
+Stage totals:
+- Pair removed by pos dedup: 145,550 (1.68%)
+- Hard-negatives removed by pos dedup: 2,671 (0.54%)
+- Overall removed by pos dedup: 148,221 (1.62%)
 
 ## Quick Usage
 
@@ -342,4 +312,9 @@ Within each subset split, rows are grouped by exact `query` string.
 - Keep the first occurrence
 - Drop all later duplicates
 
+5. Positive-document (`pos`) deduplication
+Within each subset split, rows are grouped by exact `pos` string.
+- Keep the first occurrence
+- Drop all later duplicates
+
 This filtering is purely rule-based (no model scoring), targeting high-noise templates and low-information positives while preserving broad retrieval coverage.