GenesisL1 ForestDataset
For regression/binary/multiclass, choose one label (target) column. For multilabel, choose multiple label columns. Rows with missing / non-numeric feature values are dropped (and for multilabel, rows with invalid label values are dropped too).
Regression predicts a number. Binary classification predicts the probability of the positive class.
Multiclass classification predicts one of the selected classes (argmax). Multilabel classification predicts an independent probability per label column.
Feature columns
Toggle which columns are used as features (X). Selected label column(s) are excluded automatically.
Upload a CSV to begin.
3D Distribution
PCA
Visualize your dataset in 3D. Default is a PCA(3) embedding over the currently selected feature columns.
Click a point to set the dataset row index for local preview.
Upload a CSV and select features to render.
Training
Core hyperparameters
More boosting rounds can improve performance but increases on-chain size. For multiclass/multilabel this is trees per class/label (total trees = rounds × K).
Deeper trees fit more complexity but grow exponentially in size.
Lower = slower learning but often better generalization.
Higher = smoother model; lower = fits noise more easily.
More bins = more candidate split thresholds (slower training). Does not change on-chain model size.
Quantile binning often improves splits for skewed features (slower to compute; uses up to 50,000 train rows).
Learning rate schedule (optional)
Schedules apply per boosting round. For multiclass/multilabel, each round trains one tree per class/label (so the schedule step affects all classes/labels in that round).
Data split & early stopping
Controls train/val/test shuffle (based on your split %) and tie-breaking.
Train %
Val %
Test 10%
Default 70/20/10. Test set is never used for training (only evaluation).
Stop when validation metric stops improving (MSE for regression; LogLoss for classification)
How many boosting rounds to wait for an improvement in the validation metric before stopping.
After model selection, retrain on Train+Val for the chosen number of boosting rounds (keeps on-chain size the same).
On-chain size estimate
Training curve
Training metric over boosting rounds (MSE for regression; LogLoss for classification). For multiclass/multilabel, one round = one tree per class/label.
Metrics
Heuristic hyperparameter search Optional
Max rounds
Runs N training rounds and picks the best validation score (lower is better). The best model will be kept for deployment.
Idle
Local preview
Train a model to enable.
Training curve
Same curve as Training tab (best round if heuristic search is enabled).
Best model metrics
Best model hyperparameters
These are the exact settings used to train the current best model (after search/refit if enabled).
Load a dataset row into inputs, then predict and compare vs the real label.
Mint Model NFT
Metadata (required)
Stored fully on-chain inside the NFT. Must be exactly 128×128 px.
Owner API access key (no wallet)
This suite grants a perpetual on-chain access key to the model owner at mint time.
Save the private key now — it cannot be recovered later.
I saved this private key
Needed for the read-only API key inference system.
Inference pricing (owner-controlled)
UI clamps 0.001–1 L1 for sanity (you can change later as owner).
Legal (required)
Loading license…
Loading terms…
Deploy
Train a model to see deploy estimate (transactions + on-chain deploy value).
Need: connect wallet, train model, generate+save owner key, name + description, 128×128 icon, accept Terms + License