Update README.md

README.md

@@ -1191,4 +1191,15 @@ modalities:
 
 # Assessors For Regression: Loss Analysis - Instance Level Results
 
-AFRLA - Instance Level Results is a collection of predictions at the instance level for eleven different regression tasks tested on 255 different tree-based models. The aim of this dataset is to provide example-level results to train assessor models to predict performance of the tree-based models.
+AFRLA - Instance Level Results is a collection of predictions at the instance level for eleven different regression tasks tested on 255 different tree-based models. The aim of this dataset is to provide example-level results to train assessor models to predict performance of the tree-based models.
+
+The dataset presents eleven sections (one per regression task), with varying degrees of performance, difficulty and characteristics from the original tasks. Every one of the 255 models was trained on a subset of the dataset used for every task, and the result shown here are the test (never-before-seen by the models) predictions. Each subset has:
+
+- An **instance identifier** indicating the instance nº from the test set. This is just an identifier and it is not usually employed for training assessors, although in some occasions it may be useful.
+- The **original task features**, the features used by the models to be trained. Along with the instance identifier, they fully describe a test example.
+- The **model features**, descriptors of the 255 models. Mainly:
+  - The model used (XGBoost, Random Forest, Decision Tree...)
+  - Hyperparameters such as the maximum depth, number of estimators if applicable...
+  - Profiling metrics such as training time, inference time or memory usage
+
+  These metrics are not recorded per example, but rather per model (that is, if the inference time is 1.2 ms, the model predicted *the entirety* of the test dataset in that time, instead of just that example), and are then casted for each example. As such, they fully describre a model.