Upload folder using huggingface_hub

metadata/README.md

@@ -0,0 +1,24 @@
+# Business Entity Resolution Dataset - Embedding Friendly Scenario
+
+Scenario: `data-splink`
+
+This synthetic dataset is designed for Vietnamese text-based Entity Resolution / Record Linkage research.
+It intentionally weakens exact identifiers and creates cross-source textual variation so that language embeddings can contribute beyond traditional fuzzy matching.
+
+## Design choices
+
+- High `tax_code` missing/noisy rate.
+- Cross-source schema mismatch.
+- Vietnamese text variants: missing diacritics, token merge, token order swap, legal type abbreviation, English translation, short-name-only variants.
+- Hard negatives: similar names, same address, same representative, same short name across different true entities.
+
+## Important files
+
+- `canonical/canonical_business_entity.csv`: synthetic clean entity layer.
+- `observed/*.csv`: raw source tables with different schemas.
+- `evaluation/ground_truth_entity_map.csv`: record-to-entity ground truth.
+- `evaluation/evaluation_labels.csv`: pair-level labels for model evaluation.
+- `metadata/dataset_summary.json`: quality summary.
+
+This notebook exports raw source tables only. It does not create a linkage view in advance.
+The linkage view should be created later inside the matching pipeline notebook.

metadata/dataset_summary.json

@@ -1,25 +1,30 @@
 {
+  "scenario_name": "data-splink",
   "n_canonical_entities": 8000,
   "n_observed_records": 20000,
+  "n_evaluation_labels": 19688,
   "records_by_source": {
-    "business_registry": 4719,
-    "inspection_record": 5488,
-    "license_record": 5130,
-    "tax_registry": 4663
+    "business_registry": 4613,
+    "inspection_record": 5305,
+    "license_record": 5395,
+    "tax_registry": 4687
   },
   "entities_by_cluster_size": {
-    "1": 1461,
-    "2": 2807,
-    "3": 2347,
-    "4": 1094,
-    "5": 242,
-    "6": 45,
-    "7": 4
-  },
-  "duplicate_entity_rate": 0.8174,
+    "1": 2511,
+    "2": 1994,
+    "3": 1684,
+    "4": 1035,
+    "5": 486,
+    "6": 195,
+    "7": 67,
+    "8": 14,
+    "9": 12,
+    "10": 2
+  },
+  "duplicate_entity_rate": 0.686125,
   "missingness_by_field": {
-    "tax_code": 0.2351,
-    "ward": 0.0229,
+    "tax_code": 0.5176,
+    "ward": 0.03805,
     "representative_name": 0.0,
     "business_name": 0.0,
     "address": 0.0,
@@ -28,29 +33,49 @@
     "district": 0.0
   },
   "top_error_types": {
-    "tax_code_missing": 4703,
-    "remove_diacritics": 4013,
-    "representative_abbreviation": 3152,
-    "address_shorten": 2786,
-    "brand_name_only": 2504,
-    "legal_type_abbreviation": 2222,
-    "free_text_address": 2221,
-    "short_name_only": 2154,
-    "industry_abbreviation": 1930,
-    "case_variation": 1869,
-    "english_translation": 1580,
-    "legal_type_translation": 1580,
-    "typo": 1447,
-    "address_format_variation": 1332,
-    "minor_address_missing": 458
-  },
-  "hard_positive_count": 6310,
-  "hard_negative_entity_count": 2000,
-  "n_evaluation_labels": 19999,
-  "evaluation_labels_by_pair_type": {
-    "random_negative": 10319,
-    "positive": 4079,
-    "hard_negative": 3399,
-    "hard_positive": 2202
-  }
+    "tax_code_missing": 10352,
+    "embedding_only_win_candidate": 6857,
+    "address_shorten": 6688,
+    "remove_diacritics": 6318,
+    "merge_tokens": 3343,
+    "english_translation": 2598,
+    "token_order_swap": 2342,
+    "legal_type_abbreviation": 2314,
+    "short_name_or_abbrev": 2278,
+    "typo": 1902,
+    "representative_abbreviation": 1901,
+    "representative_remove_diacritics": 1640,
+    "representative_token_order_swap": 1264,
+    "industry_abbreviation": 1095,
+    "tax_code_typo": 1034,
+    "address_remove_diacritics": 1005,
+    "address_drop_admin": 761,
+    "tax_code_wrong": 696,
+    "legal_type_abbreviation_light": 551
+  },
+  "variant_difficulty_distribution": {
+    "easy": 0.18175,
+    "hard": 0.44605,
+    "medium": 0.3722
+  },
+  "evaluation_pair_type_distribution": {
+    "positive_hard": 6000,
+    "negative_random": 5999,
+    "negative_hard": 3849,
+    "positive_regular": 3840
+  },
+  "evaluation_pair_difficulty_distribution": {
+    "hard": 14757,
+    "medium": 4432,
+    "easy": 499
+  },
+  "hard_negative_distribution": {
+    "same_name_different_legal_type": 1159,
+    "similar_name_same_industry": 1020,
+    "same_short_name_same_district": 1003,
+    "same_address_different_business": 1000,
+    "same_representative_different_business": 968,
+    "same_short_name_different_province": 871
+  },
+  "design_note": "Dataset v2 keeps duplicate rate more realistic, weakens tax_code, increases semantic/text variants, and adds hard negatives so embedding has a meaningful role."
 }