module-attribute
#
The type of a column in a datasource
module-attribute
#
JobStatus = Literal[
"INITIALIZED",
"DISPATCHED",
"WAITING",
"PROCESSING",
"COMPLETED",
"FAILED",
"ABORTING",
"ABORTED",
]
Status of job in the job queue
module-attribute
#
Status of worker in the worker pool
Bases: TypedDict
instance-attribute
#
Which loss family to train with. ‘prediction’ adds a linear head on top of embeddings. ‘contrastive’ and ‘triplet’ train embeddings directly for similarity. ‘proxy’ uses class-proxy similarity. Each loss has different defaults for batch size, evaluation, and learning rate.
instance-attribute
#
Whether the model predicts discrete labels or continuous scores. The default is classification; set regression when training against a score column.
instance-attribute
#
Peak learning rate after warmup. Higher values train faster but risk instability. Tuple searches log-uniformly between (min, max).
instance-attribute
#
Total samples per training step. Larger batches give more stable gradients and better contrastive negatives but use more memory. Automatically split across GPUs and gradient accumulation steps. List to try multiple sizes in a sweep.
instance-attribute
#
Maximum number of training steps. Overrides epochs when set. Useful
for quick validation runs or capping long-running jobs.
instance-attribute
#
Learning rate warmup. int = number of steps, float = fraction of total steps (0–1). Warmup helps stabilize early training. Tuples/lists follow the same int/float convention.
instance-attribute
#
L2 regularization strength. Helps prevent overfitting. Typical range: 0.0 to 0.1.
instance-attribute
#
How the learning rate changes after warmup. ‘linear’ decays to zero, ‘cosine’ follows a cosine curve, ‘constant’ stays flat. List to compare schedulers in a sweep.
instance-attribute
#
Inverse temperature (1/τ) for contrastive and proxy losses — controls how sharply the model distinguishes similar from dissimilar pairs. Higher = more discriminative. Only used by contrastive and proxy losses.
instance-attribute
#
Gaussian kernel width for contrastive regression. None auto-tunes from the training score standard deviation. Only used by contrastive regression.
instance-attribute
#
L2-normalize embeddings before the prediction head. Can improve stability when embedding magnitudes vary. Only used by prediction loss.
instance-attribute
#
Maximum token length for input text. Longer sequences are truncated. ‘max’ fits the longest sample, ‘p95’ / ‘p99’ covers 95% / 99% of samples (saves memory), or set an int for an explicit limit.
instance-attribute
#
Which end to cut when text exceeds max_seq_length. ‘right’ keeps the beginning.
instance-attribute
#
Task instruction for instruction-tuned models (e.g. ‘Classify this text’). Formatted into ‘Instruct: {instruction}\nQuery: ’ by the embedding model. None uses the model’s built-in default prompt (if any).
instance-attribute
#
Use bfloat16 mixed precision to halve memory usage and speed up training. None auto-enables on supported GPUs (A100, H100, etc.). Set False to force full precision.
instance-attribute
#
Maximum samples that fit in one GPU’s memory per forward pass. None auto-estimates from your GPU’s memory and model size. Override if you hit out-of-memory errors or want tighter control.
instance-attribute
#
Trade compute for memory by recomputing activations during backward. Roughly halves memory at ~30% slower training. None auto-enables when it would avoid quality-degrading workarounds (mini-batching for contrastive, or fitting an otherwise impossible triplet batch).
instance-attribute
#
Share contrastive negatives across all GPUs. Improves quality when the batch is too large for a single GPU. None auto-enables when needed. Only relevant for contrastive loss with multiple GPUs.
instance-attribute
#
How to measure model quality during training. ‘head’ reuses the prediction head (fast, prediction only). ‘neighbor’ runs nearest-neighbor search (works for all losses, slower). ‘loss’ uses the training loss on held-out data (cheapest). None picks ‘head’ for prediction, ‘neighbor’ otherwise.
instance-attribute
#
How often to evaluate. int = every N training steps, ‘epoch’ = once per epoch, ‘end’ = only after training finishes, ‘off’ = skip evaluation entirely. None picks a sensible default based on eval_method.
instance-attribute
#
Batch size for evaluation inference. None auto-detects from GPU memory.
instance-attribute
#
Cap on auto-detected eval_batch_size. Lower this if evaluation hits OOM.
instance-attribute
#
Number of nearest neighbors to consider for neighbor evaluation.
instance-attribute
#
Reduce the neighbor search pool for faster evaluation. int = use this many train samples, float = use this fraction. None uses the full train set.
instance-attribute
#
Stop training when the eval metric stops improving. True = stop after 2 evaluations without improvement, int = custom patience count, False = always train for all epochs. Requires eval_steps to run during training (not ‘end’ or ‘off’).
instance-attribute
#
Minimum improvement to count as progress for early stopping. 0.0 means any improvement resets the patience counter.
instance-attribute
#
Number of hyperparameter configurations to try. 1 = single training run. Values > 1 activate sweep mode, which uses Optuna to search over any parameter specified as a range or list.
instance-attribute
#
Number of random trials before the optimizer starts making informed suggestions. 0 = start optimizing immediately (good for small trial budgets).
instance-attribute
#
Random seed for reproducibility. Controls data shuffling, dropout, weight initialization, and sweep trial sampling.
instance-attribute
#
Print training metrics (loss, learning rate, etc.) every N steps.
instance-attribute
#
Advanced HuggingFace Accelerator settings. Most users can leave this as None.
instance-attribute
#
Additional HuggingFace TrainingArguments not exposed above. Merged directly into the training configuration.
Bases: TypedDict
instance-attribute
#
Minimum uniformity score (0-1) required for a memory to be considered as a representative. A uniformity of 1.0 means all neighbors are the same class. Lower values allow more flexibility but may select fewer prototypical examples. Default is 1.0.
Bases: TypedDict
instance-attribute
#
Explained variance score of the predictions
instance-attribute
#
Mean of anomaly scores across the dataset
instance-attribute
#
Median of anomaly scores across the dataset
instance-attribute
#
Variance of anomaly scores across the dataset
Bases: TypedDict
instance-attribute
#
Mean of anomaly scores across the dataset
instance-attribute
#
Median of anomaly scores across the dataset
instance-attribute
#
Variance of anomaly scores across the dataset
instance-attribute
#
Average precision (area under the curve of the precision-recall curve)
instance-attribute
#
Confusion matrix where the entry at row i, column j is the count of samples with true label i predicted as label j