Model Playground

Pattern

ROC Distribution — All Runs

Select a pattern row. The chosen saved run is only for inspection; training uses the inputs below unless you load params from a run.

Pattern

Labeled

Saved runs

Promoted

Asymmetric Choice

Confirm Shaming

Misdirection

Privacy Zuckering

Scarcity

Social Proof

Trick Question

Urgency

Loader Sampling + split controls.

min_pos

Minimum number of positive (dark) examples required to run training. If fewer are available, the run should fail fast.

min_neg

Minimum number of negative (not-dark) examples to include. Helps stabilize training when positives are scarce.

neg_ratio

Target negative-to-positive ratio when sampling negatives (e.g., 60 means ~60 negatives per positive), capped by max_neg.

max_neg

Upper cap on the number of negative examples pulled in. Prevents runs from getting huge when the corpus is large.

seed

Random seed for negative sampling / dataset construction. Change this to see how sensitive results are to which negatives you drew.

split_seed

Seed for the grouped stratified split (StratifiedGroupKFold shuffle). Controls which groups land in train vs test.

n_splits

Number of folds used by StratifiedGroupKFold. The playground uses the first split as train/test.

topk_importance

How many features to return in the “Top Importance” chart (by gain). UI-only; it doesn’t affect training.

XGBoost Model hyperparameters.

objective

Locked for consistency.

eval_metric

Stable loss for imbalanced binary.

max_depth

Lower if you see overfit.

eta (learning rate)

0.05 is conservative.

num_boost_round

# trees.

scale_pos_weight

Auto from imbalance.

subsample colsample_bytree

lambda (L2) seed