v4.6.1: audited real data, NLEmbedding 512-dim; remove defective MiniLM v4.5.0

BelgoClassifier.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5a51c134704e9c636ef13691c26046c6c901175fd3eb582d561a397a26ef7f7b
-size 4035
+oid sha256:f913ed77d7e3f937616842fa484f1380ecbf1fe959fcc8dda5691315a982a9d5
+size 7353

BelgoClassifier.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f26080c13cf52a597d0a9688b2d7286632970c5783b7bc2bfa31146cf4dd1a7
-size 272132
+oid sha256:80854ba9f26f940669d959c94e98339b94cb9595de9455977090cbe9e830cf43
+size 869764

BelgoClassifier.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "CD82C1A0-5651-4C22-8E88-9DBE6295C60C": {
+        "4588435D-4AEC-4B65-A3DF-1DAEBC543087": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Weights",
             "name": "weights",
             "path": "com.apple.CoreML/weights"
         },
-        "FF3075AB-1A30-4FAB-9ABA-AD94DB2C6FF9": {
+        "80326F21-FE20-4042-A4D6-CF9B9607C139": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
         }
     },
-    "rootModelIdentifier": "FF3075AB-1A30-4FAB-9ABA-AD94DB2C6FF9"
+    "rootModelIdentifier": "80326F21-FE20-4042-A4D6-CF9B9607C139"
 }

BelgoClassifier_v4.5.0.mlmodelc.zip → BelgoClassifier_v4.6.1.mlmodelc.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0a2fb8fdc4c09a2c824eccb6fc0ccae639535b85727a0c14374c71239b87f045
-size 254212
+oid sha256:b9bbb7b3f0b5c2911c7bd853143e475b0320ded8b166bdcb4bd561da62214fe6
+size 803118

README.md

@@ -1,76 +1,47 @@
 ---
 license: mit
 tags:
-- coreml
-- transaction-classification
-- belgian-finance
-- ios
-language:
-- nl
-- fr
-- en
+  - coreml
+  - text-classification
+  - finance
+  - belgium
 ---
 
-# BelgoClassifier - Belgian Transaction Categorizer
+# BelgoClassifier — Belgian bank-transaction categorizer (Core ML)
 
-CoreML model for categorizing Belgian bank transactions into 15 categories.
+**⚠️ Archive channel.** This repository is a public archive of the Stay Steady
+app's on-device transaction classifier. The app itself distributes models via
+its own backend — files here are for reference and research, not consumed by
+any released app.
 
-## Model Details
+## Current model: v4.6.1 (2026-07-20)
 
-- **Version:** 4.5.0
-- **Accuracy:** 86.7%
-- **Framework:** CoreML (iOS/macOS)
-- **Input:** 384-dimensional embedding vector
-- **Output:** 15-class probability distribution
+`BelgoClassifier_v4.6.1.mlmodelc.zip` — classifier head over **Apple
+NLEmbedding sentence embeddings (512-dim, English)**. 34 spending categories.
+Trained on 21,565 samples: synthetic Belgian transaction text plus manually
+audited real bank transactions (personal/P2P counterparties excluded, merchant
+names canonicalized — branch towns, store numbers and card fragments stripped).
+Training accuracy 99.38%.
 
-## Categories
+**Runtime contract:** embed input text with
+`NLEmbedding.sentenceEmbedding(for: .english)` (512-dim), UPPERCASE,
+diacritics folded, merchant canonicalized. Input tensor `embeddings`
+`[1, 512] Float32`; output `probabilities` over 34 classes.
 
-1. groceries - Supermarkets, food stores
-2. restaurants - Restaurants, fast food, delivery
-3. transport - Public transport, fuel, parking
-4. utilities - Electricity, gas, water
-5. telecom - Phone, internet providers
-6. healthcare - Medical, pharmacy, insurance
-7. insurance - All insurance types
-8. housing - Rent, mortgage
-9. entertainment - Cinema, events, sports
-10. shopping - Retail, online shopping
-11. subscriptions - Streaming, software
-12. income - Salary, refunds
-13. transfers - Bank transfers
-14. cash - ATM withdrawals
-15. other - Uncategorized
+## ⚠️ v4.5.0 is defective — do not use
 
-## Usage in iOS
+`BelgoClassifier_v4.5.0.mlmodelc.zip` has been **removed** from this
+repository. It was trained by a different pipeline against
+`all-MiniLM-L6-v2` embeddings (**384-dim**) and therefore cannot score
+NLEmbedding vectors at all; it also measured only 86.7% validation accuracy
+and carried a merchant+location labeling bias. If you have a cached copy,
+discard it.
 
-The model expects a 384-dimensional embedding vector (from sentence-transformers/all-MiniLM-L6-v2 or Apple's NLEmbedding).
+## Version history
 
-```swift
-// Load model
-let model = try MLModel(contentsOf: modelURL)
-
-// Create input
-let embedding = getEmbedding(for: transactionText) // 384-dim vector
-let input = try MLDictionaryFeatureProvider(dictionary: ["embeddings": embedding])
-
-// Predict
-let output = try model.prediction(from: input)
-let probabilities = output.featureValue(for: "probabilities")
-```
-
-## Training
-
-Trained on synthetic Belgian transaction data including:
-- Belgian supermarkets (Colruyt, Delhaize, Carrefour, etc.)
-- Belgian banks and insurers
-- Belgian telecom providers
-- Common Belgian merchants
-
-## License
-
-MIT License - Free to use in commercial applications.
-
-## Links
-
-- App: BelgoBudgetto (iOS)
-- Training code: Private repository
+| Version | Embeddings | Notes |
+|---|---|---|
+| 4.6.1 | NLEmbedding 512 | Audited real data, canonicalized merchants — current |
+| ~~4.5.0~~ | MiniLM 384 | **Defective, removed** (wrong pipeline, 86.7%) |
+| 2.2.0 | NLEmbedding 512 | First NLEmbedding retrain |
+| 2.1.0 / 1.0.0 | — | Early iterations, archive only |

model_manifest.json

@@ -1,9 +1,9 @@
 {
-  "latest_version": "4.5.0",
+  "latest_version": "4.6.1",
   "min_app_version": "1.0.0",
-  "released_at": "2026-06-28",
-  "accuracy": 0.867,
-  "download_url": "https://huggingface.co/githubDCS/belgo-classifier/resolve/main/BelgoClassifier_v4.5.0.mlmodelc.zip",
-  "size_mb": 0.24,
-  "changelog": "Retrained on 32 community merchant corrections (consensus gate removed; merchant-only filter keeps savings, drops person/peer counterparties). Val accuracy 86.7% (34 classes)."
+  "released_at": "2026-07-20",
+  "accuracy": 0.9938,
+  "download_url": "https://huggingface.co/githubDCS/belgo-classifier/resolve/main/BelgoClassifier_v4.6.1.mlmodelc.zip",
+  "size_mb": 0.77,
+  "changelog": "Retrained on audited real Belgian transactions (1221-row label audit, Fortis PDF data, merchant canonicalization). NLEmbedding 512-dim. Supersedes the defective v4.5.0 MiniLM artifact."
 }