Upload README.md with huggingface_hub

README.md

@@ -16,25 +16,22 @@ size_categories:
 
 # US Bank Transaction Categories v2 — Synthetic Dataset
 
-**Successor to [v1](https://huggingface.co/datasets/DoDataThings/us-bank-transaction-categories).** 24,000 sign-prefixed transaction descriptions with expanded merchant coverage, PayPal wrapper patterns, and a refined 16-category taxonomy.
+68,000 sign-prefixed transaction descriptions across 17 spending categories, modeled after real US bank statement formats. Designed for training classifiers that work on actual bank data — not the clean `"Starbucks coffee"` descriptions that most datasets use.
 
-## What Changed from v1
+**Successor to [v1](https://huggingface.co/datasets/DoDataThings/us-bank-transaction-categories).**
 
-| | v1 | v2 |
-|---|---|---|
-| **Format** | `description, category` | `[debit]/[credit] description, category` |
-| **Categories** | 16 (includes Housing) | 16 (Housing split into Rent + Mortgage removed) |
-| **Samples** | 16,000 (1,000/category) | 24,000 (1,500/category) |
-| **Merchants** | ~300 | ~500+ |
-| **PayPal wrappers** | Few categories | All spending categories (8% of samples) |
-| **POS prefixes** | SQ* | SQ*, TST* (Toast), CLV* (Clover) |
-| **Transfer patterns** | Basic P2P + CC payments | + brokerage sweeps, fintech, wire, cashier's checks, ATM, crypto, BNPL |
+## Why This Dataset
+
+Real bank transaction data is private. But the _formats_ are universal — Chase, Apple Card, PayPal, Capital One, Mercury all produce descriptions with predictable structures. This dataset captures those structures with 500+ real merchant names, randomized store numbers, addresses, and reference codes.
 
-## Why Sign Prefixes?
+### Why Sign Prefixes
 
-v1 couldn't distinguish Income from Transfer because the description alone is ambiguous. "VENMO CASHOUT" as a credit (money arriving) is Income. As a debit (money leaving) it's Transfer. The `[debit]`/`[credit]` prefix encodes the transaction direction from the cardholder's perspective.
+Bank transaction descriptions are ambiguous without direction. "VENMO CASHOUT" as a credit is income — as a debit it's a transfer. `[debit]`/`[credit]` encodes the cardholder's perspective so the model can disambiguate.
 
-Sign distributions per category reflect real-world patterns — spending categories are 95%+ debit, Income is 97% credit, Transfer is 50/50.
+Sign distributions per category reflect real-world patterns:
+- Spending categories: 95%+ debit
+- Income: 97% credit
+- Transfer: 50/50 (bidirectional by nature)
 
 ## Format
 
@@ -43,44 +40,62 @@ CSV with two columns:
 ```
 description,category
 "[debit] AMAZON MKTPL*K8R2M5VN7","Shopping"
-"[credit] Zelle payment from JAMES GARCIA COFZ4RFI1WBQ","Income"
-"[debit] PAYPAL           INST XFER  NETFLIX         WEB ID: PAYPALSI42","Subscription"
-"[debit] SCHWAB MONEYLINK PPD ID: 207802622","Transfer"
+"[credit] Patelco CU       PAYROLL                    PPD ID: 321076470","Income"
+"[debit] PreApproved Payment Bill User Payment: Netflix","Subscription"
+"[debit] SCHWAB BROKERAGE MONEYLINK PPD ID: 9005586224","Transfer"
 "[credit] AUTOMATIC PAYMENT - THANK","Transfer"
-"[debit] PreApproved Payment Bill User Payment: Google","Subscription"
+"[debit] PP*SAFEWAY","Groceries"
+"[debit] TST*GOLDEN DRAGON - OAKLAND","Restaurants"
 ```
 
-## Categories (16)
+## Categories (17)
 
 | Category | Count | Examples |
 |----------|-------|----------|
-| Restaurants | 1,500 | `[debit] MCDONALD'S F28471`, `[debit] TST*SAKURA SUSHI`, `[debit] CLV*BURGER JOINT` |
-| Groceries | 1,500 | `[debit] SAFEWAY #1197`, `[debit] WHOLEFDS FRE #10467`, `[debit] 365 MARKET` |
-| Shopping | 1,500 | `[debit] Amazon.com*K8R2M5VN7`, `[debit] TIKTOK SHOP`, `[debit] TOTAL WINE` |
-| Transportation | 1,500 | `[debit] CHEVRON 0385291`, `[debit] CHARGEPOINT *STATION 1234`, `[debit] CA DMV FEE` |
-| Entertainment | 1,500 | `[debit] VALVE STEAM PURCHASE`, `[debit] DRAFTKINGS SPORTSBOOK`, `[debit] AMC THEATRES` |
-| Utilities | 1,500 | `[debit] PG&E`, `[debit] REPUBLIC SERVICES TRASH`, `[debit] SUNRUN SOLAR` |
-| Subscription | 1,500 | `[debit] PreApproved Payment Bill User Payment: Microsoft`, `[debit] OBSIDIAN.MD`, `[debit] RAPIDAPI` |
-| Healthcare | 1,500 | `[debit] TELADOC TELEHEALTH`, `[debit] ASPEN DENTAL`, `[debit] LENSCRAFTERS` |
-| Insurance | 1,500 | `[debit] FARMERS INS BILLING`, `[debit] STATE FARM PREMIUM`, `[debit] GEICO` |
-| Rent | 1,500 | `[debit] EQUITY RESIDENTIAL RENT PAYMENT`, `[debit] GREYSTAR LEASE PAYMENT` |
-| Travel | 1,500 | `[debit] ROYAL CARIBBEAN`, `[debit] TSA PRECHECK`, `[debit] MARRIOTT` |
-| Education | 1,500 | `[debit] WYZANT TUTORING`, `[debit] COURSERA`, `[debit] UNIVERSITY OF TUITION` |
-| Personal Care | 1,500 | `[debit] SPORT CLIPS`, `[debit] SEPHORA`, `[debit] MASSAGE ENVY` |
-| Transfer | 1,500 | `[debit] COINBASE ACH TRANSFER`, `[debit] WIRE TRANSFER TO NAME`, `[credit] ATM DEPOSIT` |
-| Income | 1,500 | `[credit] EMPLOYER PAYROLL PPD ID:`, `[credit] SSA TREAS 310 FED SAL`, `[credit] DOORDASH DASHERPAY` |
-| Fees | 1,500 | `[debit] OVERDRAFT FEE`, `[debit] ATM SURCHARGE`, `[debit] PAPER STATEMENT FEE` |
-
-## Eight Format Templates
-
-1. **ACH (Chase checking)** — `INSTITUTION PURPOSE DETAILS WEB/PPD ID: CODE`
-2. **Merchant + ID (Chase credit)** — `MERCHANT #1234` or `MERCHANT*ORDERID`
-3. **Full address (Apple Card)** — `MERCHANT ADDRESS CITY ZIP STATE COUNTRY`
-4. **PayPal wrappers** — `PreApproved Payment Bill User Payment: MERCHANT`, `Express Checkout Payment: MERCHANT`, `PP*MERCHANT`, `PAYPAL *MERCHANT`, `PAYPAL INST XFER MERCHANT WEB ID:`
-5. **Action prefix (Capital One)** — `Withdrawal from DESCRIPTION` / `Deposit from DESCRIPTION`
-6. **POS prefixes** — `SQ *MERCHANT`, `TST*MERCHANT`, `CLV*MERCHANT`
-7. **Compressed names** — `DAVESHOTCHICKEN`, `CHICKFILA`, `WHOLEFDS FRE`
-8. **Simple** — `MERCHANT` or `MERCHANT.COM`
+| Restaurants | 4,000 | `TST*SAKURA SUSHI`, `DAVESHOTCHICKEN`, `CLV*BURGER JOINT`, `PP*CHIPOTLE` |
+| Groceries | 4,000 | `SAFEWAY #1197`, `WHOLEFDS FRE #10467`, `PYPL*TRADER JOE'S`, `365 MARKET` |
+| Shopping | 4,000 | `Amazon.com*K8R2M5VN7`, `TIKTOK SHOP`, `Express Checkout Payment: TARGET`, `TOTAL WINE` |
+| Transportation | 4,000 | `CHEVRON 0385291`, `CHARGEPOINT *STATION 1234`, `COSTCO GAS #1061`, `CA DMV FEE` |
+| Entertainment | 4,000 | `VALVE STEAM PURCHASE`, `DRAFTKINGS SPORTSBOOK`, `PreApproved Payment: Valve Corp.` |
+| Utilities | 4,000 | `PG&E`, `PGANDE WEB ONLINE`, `REPUBLIC SERVICES TRASH`, `SUNRUN SOLAR` |
+| Subscription | 4,000 | `CURSOR USAGE`, `X CORP. PAID FEATURES`, `PYPL*NETFLIX`, `SALESFORCE; Billing` |
+| Healthcare | 4,000 | `TELADOC TELEHEALTH`, `KAISER PERMANENTE`, `PP*ASPEN DENTAL`, `BETTERHELP` |
+| Insurance | 4,000 | `FARMERS INS BILLING`, `HOMESERVE USA`, `PreApproved Payment: STATE FARM` |
+| Mortgage | 4,000 | `ROCKET MORTGAGE`, `PATELCO CU MORTGAGE`, `SOFI MORTGAGE PAYMENT`, `Principal Pmt` |
+| Rent | 4,000 | `EQUITY RESIDENTIAL RENT PAYMENT`, `GREYSTAR LEASE PAYMENT` |
+| Travel | 4,000 | `ROYAL CARIBBEAN`, `TSA PRECHECK`, `SWA INFLIGHT WIFI`, `MARRIOTT` |
+| Education | 4,000 | `COURSERA`, `Express Checkout Payment: Scholastic Inc`, `BRILLIANT.ORG` |
+| Personal Care | 4,000 | `SPORT CLIPS`, `SEPHORA`, `360 FITNESS LLC`, `TST*ISLAND VINTAGE SHAVE` |
+| Transfer | 4,000 | `COINBASE ACH TRANSFER`, `WIRE TRANSFER TO NAME`, `ATM DEPOSIT`, `DDA TO DDA` |
+| Income | 4,000 | `PAYROLL PPD ID:`, `SSA TREAS 310 FED SAL`, `DOORDASH DASHERPAY` |
+| Fees | 4,000 | `OVERDRAFT FEE`, `ATM SURCHARGE`, `PAPER STATEMENT FEE` |
+
+## Eight Bank Statement Formats
+
+Every spending category produces descriptions in all major US bank formats:
+
+| Format | Structure | Banks |
+|---|---|---|
+| **Chase ACH** | `INSTITUTION     PURPOSE        PPD/WEB ID: CODE` | Chase checking |
+| **Chase merchant** | `MERCHANT #STORE` or `MERCHANT*ORDERID` | Chase credit cards |
+| **Apple Card** | `MERCHANT ADDRESS CITY ZIP STATE COUNTRY` | Apple Card |
+| **PayPal** | `PreApproved Payment: MERCHANT`, `PP*`, `PYPL*`, `PAYPAL *`, `INST XFER` | PayPal (as card issuer) |
+| **Capital One** | `Withdrawal from MERCHANT`, `Preauthorized Deposit from MERCHANT` | Capital One |
+| **Mercury** | `MERCHANT; Description` or just `MERCHANT` | Mercury, neobanks |
+| **POS** | `SQ *MERCHANT`, `TST*MERCHANT`, `CLV*MERCHANT` | Square, Toast, Clover |
+| **Simple** | `MERCHANT`, `MERCHANT.COM`, `MERCHANT INC.` | Various |
+
+### PayPal as a Bank Format
+
+PayPal isn't just a payment wrapper — it's a card issuer. People use PayPal credit/debit cards at any merchant. This dataset treats PayPal formats as first-class bank statement structures:
+
+- `PreApproved Payment Bill User Payment: STARBUCKS` → Restaurants
+- `PP*SAFEWAY` → Groceries
+- `PYPL*NETFLIX` → Subscription
+- `Express Checkout Payment: TARGET` → Shopping
+- `PAYPAL INST XFER MEDIUM.COM WEB ID: PAYPALSI77` → Subscription
+
+PayPal-formatted descriptions appear in all spending categories at realistic rates (5-15% of samples).
 
 ## Variation Dimensions
 
@@ -88,20 +103,21 @@ description,category
 - **Spacing:** normal, extra-padded (ACH style), compressed
 - **Store numbers:** `#1234`, `1234`, `#01234`, absent
 - **Order IDs:** `*ORDERID` (Amazon style)
-- **POS prefixes:** `SQ *` (Square), `TST*` (Toast), `CLV*` (Clover)
+- **POS prefixes:** `SQ *`, `TST*`, `CLV*`
+- **PayPal prefixes:** `PP*`, `PYPL*`, `PAYPAL *`, `PreApproved Payment`, `Express Checkout`
 - **Addresses:** full, partial, zip-smashed-into-city (Apple Card quirk)
+- **Compressed names:** `DAVESHOTCHICKEN`, `CHICKFILA`, `WHOLEFDS FRE`, `TRADERJOES`
 - **Cities:** 36 US cities across multiple states
-- **Sign prefixes:** `[debit]` / `[credit]` with category-appropriate distributions
-- **Abbreviated merchant names:** real bank statement abbreviations
+- **Sign prefixes:** `[debit]`/`[credit]` with category-appropriate distributions
 
 ## Design Decisions
 
-- **No "Mortgage" category.** Mortgage transactions are classified by account type, not description. Every transaction on a mortgage account is a mortgage payment by definition.
+- **No "Housing" category.** Split into Mortgage (model-classified) and Rent (model-classified). Home maintenance → Shopping.
 - **No "Business" category.** Whether a transaction is a business expense depends on the _account_, not the description.
-- **Transfer vs Income.** Sign prefix is the primary signal. `[credit] VENMO CASHOUT` = Income. `[debit] VENMO PAYMENT TO` = Transfer.
-- **Subscription vs Shopping.** `Amazon.com*ORDERID` = Shopping. `AMAZON WEB SERVICES` = Subscription. `X CORP. PAID FEATURES` = Subscription. `TIKTOK SHOP` = Shopping.
-- **PayPal wrappers across all categories.** The model learns to look past `PreApproved Payment Bill User Payment:` and classify based on the merchant inside.
-- **Balanced classes.** 1,500 samples per category prevents the classifier from defaulting to the most common class.
+- **Transfer vs Income uses sign.** `[credit] VENMO CASHOUT` = Income. `[debit] VENMO PAYMENT TO` = Transfer. The sign prefix is the primary disambiguator.
+- **Subscription vs Shopping uses format cues.** `Amazon.com*ORDERID` = Shopping. `AMAZON WEB SERVICES` = Subscription. `X CORP. PAID FEATURES` = Subscription. `TIKTOK SHOP` = Shopping.
+- **500+ real merchants** across restaurants (including culturally diverse cuisines), groceries, SaaS/AI tools, fintech platforms, crypto brokerages, EV charging, gambling/sportsbooks, and more.
+- **Balanced classes.** 4,000 samples per category prevents the classifier from defaulting to the most common class.
 
 ## Usage
 
@@ -115,14 +131,14 @@ print(ds["train"][0])
 
 ## Trained Model
 
-A DistilBERT model fine-tuned on this dataset is available at [DoDataThings/distilbert-us-transaction-classifier-v2](https://huggingface.co/DoDataThings/distilbert-us-transaction-classifier-v2).
+A DistilBERT model fine-tuned on this dataset is available at [DoDataThings/distilbert-us-transaction-classifier-v2](https://huggingface.co/DoDataThings/distilbert-us-transaction-classifier-v2) — 99.9% validation accuracy, 96% of real-world classifications at 0.90+ confidence.
 
 ## Generator
 
-The synthetic data generator script is open source:
+The synthetic data generator is open source:
 
 ```bash
-node scripts/generate-training-data.js --count 1500
+node scripts/generate-training-data.js --count 4000  # 4,000 per category
 ```
 
 Available at [github.com/wnstnb/foliome](https://github.com/wnstnb/foliome).