Publish Vietnamese business ER dataset v1.0.0

README.md

@@ -1,131 +1,28 @@
 ---
 language:
 - vi
-pretty_name: Vietnamese Business Entity Resolution Dataset — Lexical 85 / Embedding Rescue
 tags:
 - entity-resolution
 - record-linkage
-- tabular
-- synthetic
 - splink
-configs:
-- config_name: business_registry
-  default: true
-  data_files:
-  - split: train
-    path: hub/business_registry/train.csv
-  - split: validation
-    path: hub/business_registry/validation.csv
-  - split: test
-    path: hub/business_registry/test.csv
-- config_name: tax_registry
-  data_files:
-  - split: train
-    path: hub/tax_registry/train.csv
-  - split: validation
-    path: hub/tax_registry/validation.csv
-  - split: test
-    path: hub/tax_registry/test.csv
-- config_name: inspection_record
-  data_files:
-  - split: train
-    path: hub/inspection_record/train.csv
-  - split: validation
-    path: hub/inspection_record/validation.csv
-  - split: test
-    path: hub/inspection_record/test.csv
-- config_name: record_split_map
-  data_files:
-  - split: train
-    path: hub/record_split_map/train.csv
-  - split: validation
-    path: hub/record_split_map/validation.csv
-  - split: test
-    path: hub/record_split_map/test.csv
-- config_name: ground_truth_entity_map
-  data_files:
-  - split: train
-    path: hub/ground_truth_entity_map/train.csv
-  - split: validation
-    path: hub/ground_truth_entity_map/validation.csv
-  - split: test
-    path: hub/ground_truth_entity_map/test.csv
-- config_name: all_positive_pairs
-  data_files:
-  - split: train
-    path: hub/all_positive_pairs/train.csv
-  - split: validation
-    path: hub/all_positive_pairs/validation.csv
-  - split: test
-    path: hub/all_positive_pairs/test.csv
-- config_name: evaluation_pairs
-  data_files:
-  - split: train
-    path: hub/evaluation_pairs/train.csv
-  - split: validation
-    path: hub/evaluation_pairs/validation.csv
-  - split: test
-    path: hub/evaluation_pairs/test.csv
-- config_name: challenge_evaluation_pairs
-  data_files:
-  - split: train
-    path: hub/challenge_evaluation_pairs/train.csv
-  - split: validation
-    path: hub/challenge_evaluation_pairs/validation.csv
-  - split: test
-    path: hub/challenge_evaluation_pairs/test.csv
-- config_name: record_generation_metadata
-  data_files:
-  - split: train
-    path: hub/record_generation_metadata/train.csv
-  - split: validation
-    path: hub/record_generation_metadata/validation.csv
-  - split: test
-    path: hub/record_generation_metadata/test.csv
-- config_name: canonical_entity
-  data_files:
-  - split: train
-    path: hub/canonical_entity/train.csv
-  - split: validation
-    path: hub/canonical_entity/validation.csv
-  - split: test
-    path: hub/canonical_entity/test.csv
-- config_name: canonical_entity_history
-  data_files:
-  - split: train
-    path: hub/canonical_entity_history/train.csv
-  - split: validation
-    path: hub/canonical_entity_history/validation.csv
-  - split: test
-    path: hub/canonical_entity_history/test.csv
-- config_name: hard_negative_relationships
-  data_files:
-  - split: train
-    path: hub/hard_negative_relationships/train.csv
-  - split: validation
-    path: hub/hard_negative_relationships/validation.csv
-  - split: test
-    path: hub/hard_negative_relationships/test.csv
+- synthetic
+- vietnamese
+pretty_name: Vietnamese Business Entity Resolution — Embedding Lift Benchmark
 ---
 
 # Vietnamese Business Entity Resolution Dataset
 
-Synthetic three-source Vietnamese business records designed to separate lexical
-matching from multilingual semantic embedding rescue.
+Synthetic benchmark for Vietnamese business entity resolution across three sources.
 
-## Core benchmark design
+## Dataset design
 
-- Dataset version: `6.0.0-embedding-lift-controlled`
-- Canonical entities: 12,000
-- Observed records: 36,000
-- Positive pairs: 36,000
-- Designed lexical-positive share: 0.850
-- Semantic-only positive count: 5,400
-- Split by entity: train 70%, validation 15%, test 15%.
+- Entities: 12,000
+- Records: 36,000
+- Sources: business_registry, tax_registry, inspection_record
+- Split: 70% train, 15% validation, 15% test by entity
+- Positive pair mix: 70% lexical, 20% semantic-only, 10% mixed/hard
 
-`all_positive_pairs` is the blocking-recall reference. `evaluation_pairs` is the
-balanced main benchmark. `challenge_evaluation_pairs` is for robustness/error
-analysis and must not be used to choose the main threshold.
+## Access boundary
 
-Matching and `u/m` estimation may read only observed source tables, split control,
-and metadata contracts. Canonical/evaluation files contain generator truth.
+Matching code may read only `observed/`, `control/`, and non-label metadata.
+`evaluation/` and `canonical/` contain ground truth and are opened only after prediction.

canonical/canonical_entities.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f6da9ab9469a75b321649e9d11e542e57838943c0ba4ea3222db3066d5894b7
+size 605533

control/record_split_map.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f13903658e83145c62d3ddaed45d8d2d3522412de872931ca48dcd2e7edf2563
+size 225114

evaluation/all_positive_pairs.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83479c2a758dc8faf1da93df81384ddf640f6f123c93e63f337804db6bb7dc57
+size 324179

evaluation/ground_truth_entity_map.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3db99c85594a0ead708fd9c002a7769db4441af0f157aff6cdfc9108530f7ee
+size 339595

evaluation/hard_negative_pairs.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eed50e9ab0a6bae8278b357b46bcde34716e9a4e6b8f8b911233f2684abeb30
+size 40586

evaluation/record_generation_metadata.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64df1d8a6d82022743ca3980112211da670b19e71b5f815253b768b3972c934b
+size 245709

metadata/dataset_metadata.json

@@ -0,0 +1,52 @@
+{
+  "dataset_version": "1.0.0",
+  "seed": 42,
+  "n_entities": 12000,
+  "n_records": 36000,
+  "sources": [
+    "business_registry",
+    "tax_registry",
+    "inspection_record"
+  ],
+  "split_weights": {
+    "train": 0.7,
+    "validation": 0.15,
+    "test": 0.15
+  },
+  "entity_scenario_shares": {
+    "lexical": 0.55,
+    "semantic_only": 0.3,
+    "mixed_hard": 0.15
+  },
+  "positive_pair_type_shares": {
+    "positive_lexical": 0.7,
+    "positive_semantic_only": 0.2,
+    "positive_mixed_hard": 0.1
+  },
+  "semantic_subtype_weights": {
+    "name_only": 0.3,
+    "address_only": 0.15,
+    "industry_only": 0.2,
+    "full_context": 0.2,
+    "multi_field": 0.15
+  },
+  "hard_negative_relationship_share": {
+    "negative_lexical_hard": 0.015,
+    "negative_semantic_hard": 0.015,
+    "negative_same_address": 0.012,
+    "negative_parent_branch": 0.012
+  },
+  "lexical_proxy_positive_recall": 0.7020277777777778,
+  "artifact_paths": [
+    "canonical/canonical_entities.parquet",
+    "control/record_split_map.parquet",
+    "evaluation/all_positive_pairs.parquet",
+    "evaluation/ground_truth_entity_map.parquet",
+    "evaluation/hard_negative_pairs.parquet",
+    "evaluation/record_generation_metadata.parquet",
+    "observed/business_registry.parquet",
+    "observed/inspection_record.parquet",
+    "observed/observed_records.parquet",
+    "observed/tax_registry.parquet"
+  ]
+}

metadata/dataset_quality_report.csv

@@ -0,0 +1,28 @@
+metric,group,value
+total_entities,all,12000.0
+total_records,all,36000.0
+records_per_entity,all,3.0
+lexical_proxy_positive_recall,all,0.7020277777777778
+entity_split_share,train,0.6999166666666666
+entity_split_share,validation,0.15025
+entity_split_share,test,0.14983333333333335
+records_by_source,business_registry,12000.0
+records_by_source,tax_registry,12000.0
+records_by_source,inspection_record,12000.0
+missingness,business_name,0.015027777777777777
+missingness,tax_code,0.2110277777777778
+missingness,address,0.05011111111111111
+missingness,representative_name,0.16927777777777778
+missingness,industry,0.04605555555555556
+positive_pair_type_share,positive_lexical,0.7
+positive_pair_type_share,positive_semantic_only,0.2
+positive_pair_type_share,positive_mixed_hard,0.1
+semantic_subtype_share,name_only,0.3
+semantic_subtype_share,industry_only,0.2
+semantic_subtype_share,full_context,0.2
+semantic_subtype_share,address_only,0.15
+semantic_subtype_share,multi_field,0.15
+hard_negative_pair_count,negative_lexical_hard,1080.0
+hard_negative_pair_count,negative_semantic_hard,1080.0
+hard_negative_pair_count,negative_same_address,870.0
+hard_negative_pair_count,negative_parent_branch,870.0

observed/business_registry.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:123402fd7994e498eb08103c61b287122bfa5b2c49b319c67c3fe70c37087b68
+size 658304

observed/inspection_record.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c40368a7f4ed114d7d05670ee76babf4a0ab3517ad418619bbcf6f81c45599ef
+size 582817

observed/observed_records.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13fc5c64776d8c1cdf315afc42bd56a0698c1dab47140a880ab924bf2f96329c
+size 1756134

observed/tax_registry.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5da65af4aacf191c1d105cf3c9e311c0a5001a9a9195b64245c4eeff1176d641
+size 770025