zjunlp Dataset Metadata and Business Rules
Dataset Files
| File | Contents |
|---|---|
payments-readme.md |
Column definitions for payments.csv |
manual.md |
Business rules, fee factor semantics, ACI codes, account types |
fees.json |
1000 fee rules; fields: ID, card_scheme, account_type, capture_delay, monthly_fraud_level, monthly_volume, merchant_category_code, is_credit, aci, fixed_amount, rate, intracountry |
payments.csv |
Transaction data (see payments-readme.md for columns) |
merchant_data.json |
Per-merchant: merchant_category_code, account_type, capture_delay, acquirers |
merchant_category_codes.csv |
MCC codes and descriptions |
acquirer_countries.csv |
country_code of acquirers |
Fee Formula
fee = fixed_amount + rate * transaction_value / 10000
null in any fee rule field means the rule applies to all values of that field.
Step-by-Step Approach
Step 1: Read the authoritative sources first
Always start by reading the relevant documentation:
with open('path/to/manual.md', 'r') as f:
manual = f.read()
with open('path/to/payments-readme.md', 'r') as f:
readme = f.read()
For column/field definitions → payments-readme.md
For fee rules and business logic → manual.md (especially Section 5)
For actual fee rule data → fees.json
Step 2: Identify question type
| Question Type | Primary Source | Strategy |
|---|---|---|
| Column name/definition | payments-readme.md | Direct lookup |
| Does concept X exist? | manual.md + fees.json + payments-readme.md | Exhaustive search |
| Boolean factor effect (True/False) | manual.md Section 5 | Use manual semantics |
| Which factors decrease → cheaper? | manual.md Section 5 | Use manual semantics |
| Volume tier values / thresholds | fees.json | Query data |
| Fee rule field values | fees.json | Query data |
| ACI code meanings | manual.md Section 4 | Direct lookup |
| Account type codes | manual.md Section 2 | Direct lookup |
Step 3: Apply business rules from manual.md
Trust the manual's explicit directional statements about fee effects, even if raw data statistics appear to contradict them (confounding variables exist in the data).
Boolean fields and their cost direction (from manual Section 5):
| Field | True means | False means | Cheaper when |
|---|---|---|---|
is_credit |
Credit transaction → typically more expensive | Debit → cheaper | False |
intracountry |
Domestic (issuer country = acquirer country) → cheaper | International → typically more expensive | True |
Numeric/categorical fields and cost direction:
| Field | Direction | Rule |
|---|---|---|
monthly_fraud_level |
Higher fraud → more expensive | Decrease → cheaper |
monthly_volume |
Higher volume → cheaper (economies of scale) | Increase → cheaper |
capture_delay |
Faster capture → more expensive | Slowing down → cheaper |
Volume tiers in fees.json: <100k, 100k-1m, 1m-5m, >5m
Fraud level tiers: <7.2%, 7.2%-7.7%, 7.7%-8.3%, >8.3%
Capture delay values: immediate, <3, 3-5, >5, manual
Step 4: For "Not Applicable" determination
If a question asks about a concept (e.g., "retry fee", "chargeback amount") that may or may not exist:
- Search manual.md for the concept
- Search fees.json fields and values
- Search payments-readme.md
- Search payments.csv column headers
Only answer Not Applicable after exhaustively confirming the concept doesn't exist in any source. The manual mentions "excessive retrying" causes downgrades but defines no monetary excessive retry fee.
Step 5: Answer formatting
Match the exact format requested and use exact strings from the source files:
- Column names: use exact case as in payments-readme.md (e.g.,
has_fraudulent_dispute, notHas_Fraudulent_Dispute) - Volume tiers: use exact strings from fees.json (e.g.,
>5m,100k-1m) - Multiple values: comma-separated list (e.g.,
monthly_fraud_level, is_credit) - Non-existent concepts:
Not Applicable
Key Business Rule Summaries
Fraud indicator column: has_fraudulent_dispute (Boolean - from payments-readme.md)
ACI values: A (Card present non-auth), B (Card present auth), C (Tokenized mobile), D (Card not present COF), E (Recurring), F (3D Secure), G (Non-3D Secure)
Account types: R (Enterprise Retail), D (Enterprise Digital), H (Enterprise Hospitality), F (Platform Franchise), S (Platform SaaS), O (Other)
Boolean factors → cheaper if True: intracountry
Boolean factors → cheaper if False: is_credit
Factors → cheaper if decreased: monthly_fraud_level, is_credit
Common Pitfall: Data vs. Manual
Raw statistical comparison of fee rates by field value can be misleading because fee rules vary across many dimensions (card_scheme, MCC, ACI, etc.). For questions about which direction a factor pushes fees, rely on the manual's explicit statements rather than naive aggregate statistics from fees.json. Use fees.json data when you need specific values, tier labels, or counts.