Document feature_engineering_data_dictionary.xlsx, outputs/ folder, and default vs feature_engineered configs

README.md

@@ -65,9 +65,29 @@ This dataset backs **Credit Risk Default**: an explainable credit-risk platform
 | Marienne Dosso | Data Scientist |
 | Bintou Ba | Data Scientist |
 
+## Dataset Splits
+
+Two versions of the same train/validation/test splits are available as separate configs:
+
+- **`default`** (used automatically if you don't pass a config name): the cleaned,
+  target-filtered, leakage-removed, imputed dataset from
+  [`notebooks/00_cleaning.ipynb`](https://github.com/BuildersLab/Credit-Risk-Default/blob/main/notebooks/00_cleaning.ipynb),
+  before feature engineering. Files under `data/before_feature_engineering/`.
+- **`feature_engineered`**: the same splits after
+  [`notebooks/02_feature_engineering.ipynb`](https://github.com/BuildersLab/Credit-Risk-Default/blob/main/notebooks/02_feature_engineering.ipynb)
+  (engineered features added, several raw columns consolidated or dropped).
+  Files under `data/after_feature_engineering/`.
+
+```python
+from datasets import load_dataset
+
+before = load_dataset("BuildersLab/loan-application-dataset")                        # default config
+after  = load_dataset("BuildersLab/loan-application-dataset", "feature_engineered")  # after feature engineering
+```
+
 ## Data Dictionary
 
-Reference files under `data_dictionary/`, produced while cleaning the raw LendingClub CSV (see [`notebooks/00_cleaning.ipynb`](https://github.com/BuildersLab/Credit-Risk-Default/blob/main/notebooks/00_cleaning.ipynb)), not Kaggle's raw column list. Each has a different schema, so each CSV is declared as its own config above (see `configs:` in the metadata) rather than being auto-merged.
+Reference files under `data_dictionary/`, produced while cleaning and engineering the LendingClub data (see [`notebooks/00_cleaning.ipynb`](https://github.com/BuildersLab/Credit-Risk-Default/blob/main/notebooks/00_cleaning.ipynb) and [`notebooks/02_feature_engineering.ipynb`](https://github.com/BuildersLab/Credit-Risk-Default/blob/main/notebooks/02_feature_engineering.ipynb)), not Kaggle's raw column list. Each has a different schema, so each CSV is declared as its own config above (see `configs:` in the metadata) rather than being auto-merged.
 
 - **`data_dictionary_summary.csv`**: one row per column: `Variable`, `Description`
   (LendingClub's official column definition), `Missing Values` (count),
@@ -83,3 +103,17 @@ Reference files under `data_dictionary/`, produced while cleaning the raw Lendin
   columns that survive the leakage/keep review (median fill, median plus a
   missing-flag column, mode plus flag, or drop rows), with the mechanism and
   reasoning behind each call.
+- **`feature_engineering_data_dictionary.xlsx`**: 4-sheet workbook covering the
+  `feature_engineered` split from `02_feature_engineering.ipynb`: a data
+  dictionary, dictionary plus per-column statistics (min/max/mean/std/percentiles
+  for numeric columns, unique/top/freq for categorical columns, missing value
+  count for all), a full feature x feature correlation matrix, and feature x
+  `default` target correlation.
+
+## Outputs
+
+Analysis artifacts produced by the notebooks, not part of the model-ready dataset
+itself, live under `outputs/`. Currently the feature x feature correlation matrix
+(`feature_correlation_matrix.csv`) and its heatmap (`feature_correlation_matrix.png`)
+from `02_feature_engineering.ipynb`; later notebooks will add artifacts like SHAP
+plots and evaluation curves here too.