Hyperparameters Reference

When you create a training, an Advanced Configuration step lets you fine-tune a curated set of hyperparameters on top of your chosen model size (Small / Medium / Large). This page documents every one of them — what it does, and which values are accepted.

Everything not listed here (model architecture, loss internals, optimizer internals, infra tuning) is fixed by the size tier you pick, so results stay comparable and reproducible across runs.

Training budget

Hyperparameter	What it controls
run_name	Identifier for the training run. Set by the platform when you create the training.
epochs	How many passes over your training data to run.
seed	Random seed — set it to get reproducible results across runs.
checkpoint_epoch_progress	How often (in fractions of an epoch) a checkpoint is saved. Keep this equal to epoch_progress under Evaluation — otherwise the best-performing checkpoint may not have a saved state to restore.

Sequence & history controls

Hyperparameter	What it controls
max_seqlen_per_client_train / max_seqlen_per_client_eval	Maximum history (number of events) read per client, for training and evaluation respectively.
min_context_train / min_context_eval	Minimum history required before a target is scored.
min_context_strict_train / min_context_strict_eval	When enabled, targets that don't meet the minimum history are excluded instead of being scored with a shorter context. Pin this per project — changing it changes which rows get scored, so metrics stop being comparable across runs.
max_seqlen_per_chunk	How much history is processed per forward pass.
seqlen_trunc_prob	Probability of randomly truncating a sequence during training — a regularizer against overfitting on very long histories.
batch_size_train / batch_size_eval	Batch size for training and evaluation. Larger eval batches don't change scores and can be slower with variable-length data.

Training dynamics

Hyperparameter	What it controls
adamw_learning_rate	Learning rate for the standard optimizer.
muon_learning_rate	Learning rate for the Muon optimizer (matrix-parameter updates). Only relevant when use_muon is on.
use_muon	Turns the Muon optimizer on/off. true | false
warmup_epoch_progress	Fraction of training spent ramping the learning rate up from its minimum before decaying.
min_adamw_learning_rate	Floor for adamw_learning_rate at the end of the schedule.
min_muon_learning_rate	Floor for muon_learning_rate at the end of the schedule.
early_stopping_eval_patience	Number of evaluations without improvement before training stops early.
threshold	Minimum improvement required for an evaluation to count as "better" for early stopping.

Loss semantics

Hyperparameter	What it controls	Options
loss_averaging	How the loss is averaged: per client (each client counts equally) or per transaction (each event counts equally).	by_client | by_transaction
loss_weighting	Whether more recent events get more weight in the loss (temporal) or all events count equally (uniform).	temporal | uniform

Evaluation & delivery

Hyperparameter	What it controls	Options
main_metric	The headline metric used to pick the best checkpoint and reported as the training's overall score.	acc | auroc | auprc | gini | precision | recall | f1 | r2 | loss
use_low_fpr_metrics	Adds a low-false-positive-rate recall metric — useful for fraud/risk use cases where false positives are costly.	true | false
epoch_progress	How often (in fractions of an epoch) evaluation runs. Keep this equal to checkpoint_epoch_progress above.
logits_output_include_target_columns	Whether the delivered score file also includes the true target values alongside the model's scores.	true | false

Data sampling

Hyperparameter	What it controls
client_sampling_ratio	Fraction of clients to use — lower it to train faster/cheaper on a subsample. 1.0 = use all clients.
client_sampling_seed	Seed controlling which clients get sampled (independent of the training seed) — change it to draw a different subsample.
stats.sampling_ratio	Fraction of data used to compute normalization statistics — a cost lever, not a quality one for most datasets.
inactivity_gap	How long without any activity counts as a gap in a client's history (e.g. 15d). Accepts a number + unit: s, m, h, d, w.

Next steps

• Set these when Creating a Training
• See how they affect results in Monitoring