Operational AI Input Specs
Fraud Detection Input Specs
Use this guide to prepare the CSV files accepted by the Fraud workspace. Training and scoring use different dataset roles: training builds the saved model version, while scoring evaluates new rows against the active model or a selected saved version.
Overview
Use this guide to prepare valid input files for this workspace. Review the required files, exact column names, example rows, and common formatting rules before uploading.
Required files
2
Optional files
0
Spec sections
2
Applies to scenario
Training Dataset
Use this CSV in Scenario when you want to train the next saved Fraud model version. The parser accepts one approved binary target alias and normalizes it to Class, plus at least one numeric or categorical feature column.
Required file
fraud-training.csv
Labeled training dataset used to produce a saved Fraud model version.
Requirements
Required columns
one target column from approved aliasesat least one feature column
Optional columns
additional numeric or categorical feature columns
Row identity
- Required row identity: Class.
- Approved target aliases are normalized to Class before training.
Approved aliases
- Class: class or label or labels or target or y or fraud or is_fraud or fraud_label or risk or risk_label or outcome
Shape requirements
- Exactly one approved target alias column is allowed in a training CSV.
Value rules and examples
- The selected target column is normalized to Class before backend training.
- Target values must be 0 or 1 for every row.
- Include at least one numeric or categorical feature column besides Class.
- Categorical feature columns are encoded by the backend.
- All-empty feature columns may be dropped before training.
Example CSV table
| V1 | V2 | V3 | Class |
|---|---|---|---|
| 0.12 | -1.31 | 2.44 | 0 |
| -0.42 | 0.87 | 1.12 | 0 |
Common mistakes
- Uploading a file that has a target alias but no usable feature columns.
- Uploading multiple target alias columns in the same training CSV.
- Using text values in the target column instead of 0 and 1.
- Assuming the scoring dataset shape can be reused for training without a target column.
Quick rules
- Every training upload needs exactly one target alias column plus at least one numeric or categorical feature column.
- Approved target aliases are Class, class, label, labels, target, y, fraud, is_fraud, fraud_label, risk, risk_label, and outcome.
- Target values must stay binary as 0 or 1.
- Categorical feature columns are encoded by the backend; all-empty columns may be dropped.
Navigate
Keep the right guide close by
Pair this upload guide with the product page for workflow context or the resources hub for related public material.
Applies to score
Scoring Dataset
Use this CSV in Score when you want to evaluate new rows. The file must align with the feature schema of the active model or the selected saved version you plan to score against.
Required file
fraud-scoring.csv
Scoring dataset used to score new records against the active model or a selected saved version.
Requirements
Required columns
every feature column required by the scoring model
Optional columns
extra columns are allowed but ignored during aligned scoring
Reference requirements
- Required scoring columns come from the active saved model's training schema.
Unsupported currently
- Class aliases are not used for scoring-schema alignment.
Value rules and examples
- Use the active model or selected saved version as the feature-schema reference.
- Every required feature column must be present with the same names used by the training schema.
- Categorical feature values are supported when that categorical column exists in the saved training schema.
- Do not include Class when preparing pure scoring rows unless your downstream process expects it separately.
Example CSV table
| V1 | V2 | V3 |
|---|---|---|
| 0.18 | -1.11 | 2.21 |
| -0.28 | 0.74 | 1.04 |
Common mistakes
- Uploading scoring rows before choosing a saved-model schema to validate against.
- Leaving out one or more required feature columns from the scoring file.
- Including text values in feature columns that the scoring schema expects to be numeric.
Quick rules
- Scoring uploads must include every numeric feature column required by the current scoring model.
- Feature names must match the active model or selected saved version schema exactly.
- Extra columns are allowed, but missing required features or non-numeric feature values will block scoring.
Navigate
Keep the right guide close by
Pair this upload guide with the product page for workflow context or the resources hub for related public material.
Navigate
Keep the right input guide close
Return to the input-spec library for another upload contract, or go back to the resource hub for quickstarts, how-to guides, and use cases.