nemo_microservices.types.beta.safe_synthesizer.safe_synthesizer_parameters_param

Module Contents

Classes

SafeSynthesizerParametersParam

API

class nemo_microservices.types.beta.safe_synthesizer.safe_synthesizer_parameters_param.SafeSynthesizerParametersParam

Bases: typing_extensions.TypedDict

data: nemo_microservices.types.beta.safe_synthesizer.data_parameters_param.DataParametersParam

None

Configuration for parameters related to how to shape or use the data being passed in.

Attributes: group_training_examples_by: Column(s) to group training examples by. order_training_examples_by: Column to order training examples by. max_sequences_per_example: Maximum number of sequences per example for training. holdout: Amount of records to holdout for evaluation. max_holdout: Maximum number of records to hold out. Overrides any behavior set by holdout parameter. random_state: Random state for holdout split to ensure reproducibility.

enable_replace_pii: bool

None

Enable replacing PII in the data.

enable_synthesis: bool

None

Enable synthesizing new data by training a model.

evaluation: nemo_microservices.types.beta.safe_synthesizer.evaluation_parameters_param.EvaluationParametersParam

None

Configuration for evaluating synthetic data quality and privacy.

This class controls which evaluation metrics are computed and how they are configured. It includes privacy attack evaluations, statistical quality metrics, and downstream machine learning performance assessments.

Attributes: enabled: Enable or disable evaluation. quasi_identifier_count: Number of quasi-identifiers to sample for privacy attacks. pii_replay_enabled: Enable PII Replay detection. pii_replay_entities: List of entities for PII Replay. If not provided, default entities will be used. pii_replay_columns: List of columns for PII Replay. If not provided, only entities will be used. mia_enabled: Enable membership inference attack evaluation for privacy assessment. aia_enabled: Enable attribute inference attack evaluation for privacy assessment. sqs_report_columns: Number of columns to include in statistical quality reports. sqs_report_rows: Number of rows to include in statistical quality reports. mandatory_columns: Columns that must be present in generated data.

generation: nemo_microservices.types.beta.safe_synthesizer.generate_parameters_param.GenerateParametersParam

None

Configuration parameters for synthetic data generation.

These parameters control how synthetic data is generated after the model is trained. They affect the quality, diversity, and validity of the generated synthetic records.

Attributes: num_records: Number of synthetic records to generate. Maximum is 130,000 records. temperature: Sampling temperature for controlling randomness (higher = more random). repetition_penalty: Penalty for token repetition (≥1.0, higher = less repetition). top_p: Nucleus sampling probability for token selection (0 < value ≤ 1). patience: Number of invalid records fraction before stopping. invalid_fraction_threshold: “The fraction of invalid records that will stop generation after the patience limit is reached.” use_structured_generation: Whether to use structured generation for better format control.

privacy: nemo_microservices.types.beta.safe_synthesizer.differential_privacy_hyperparams_param.DifferentialPrivacyHyperparamsParam

None

Hyperparameters for differential privacy during training.

These parameters configure differential privacy (DP) training using DP-SGD algorithm. When enabled, they provide formal privacy guarantees by adding calibrated noise during training.

Attributes: dp_enabled: Enable differential privacy training with DP-SGD algorithm. epsilon: Target privacy budget (ε) - lower values provide stronger privacy. delta: Probability of privacy failure (δ) - should be much smaller than 1/n where n is the number of training records. per_sample_max_grad_norm: Maximum L2 norm for gradient clipping per sample.

replace_pii: nemo_microservices.types.beta.safe_synthesizer.pii_replacer_config_param.PIIReplacerConfigParam

None

Configuration for PII replacer. Used to define how PII data should be detected and replaced in a dataset.

Attributes: globals Global configuration options. steps: List of transformation steps to perform on input data.

Methods: get_default_config: Returns a default configuration instance.

training: nemo_microservices.types.beta.safe_synthesizer.training_hyperparams_param.TrainingHyperparamsParam

None

Hyperparameters that control the training process behavior.

This class contains all the fine-tuning hyperparameters that control how the model learns, including learning rates, batch sizes, LoRA configuration, and optimization settings. These parameters directly affect training performance and quality.

Attributes: num_input_records_to_sample: Number of records the model will see during training. This parameter is a proxy for training time. For example, if its value is the same size as the input dataset, this is like training for a single epoch. If its value is larger, this is like training for multiple (possibly fractional) epochs. If its value is smaller, this is like training for a fraction of an epoch. Supports ‘auto’ where a reasonable value is chosen based on other config params and data. batch_size: The batch size per device for training. gradient_accumulation_steps: Number of update steps to accumulate the gradients for, before performing a backward/update pass. This technique increases the effective batch size that will fit into GPU memory. weight_decay: The weight decay to apply (if not zero) to all layers except all bias and LayerNorm weights in the AdamW optimizer. warmup_ratio: Ratio of total training steps used for a linear warmup from 0 to the learning rate. lr_scheduler: The scheduler type to use. See the HuggingFace documentation of SchedulerType for all possible values. learning_rate: The initial learning rate for AdamW optimizer. lora_r: The rank of the LoRA update matrices, expressed in int. Lower rank results in smaller update matrices with fewer trainable parameters. lora_alpha_over_r: The ratio of the LoRA scaling factor (alpha) to the LoRA rank. Empirically, this parameter works well when set to 0.5, 1, or 2. lora_target_modules: The list of transformer modules to apply LoRA to. Possible modules: ‘q_proj’, ‘k_proj’, ‘v_proj’, ‘o_proj’, ‘gate_proj’, ‘up_proj’, ‘down_proj’ use_unsloth: Whether to use unsloth. rope_scaling_factor: Scale the base LLM’s context length by this factor using RoPE scaling. validation_ratio: The fraction of the training data that will be used for validation. The range should be 0 to 1. If set to 0, no validation will be performed. If set larger than 0, validation loss will be computed and reported throughout training. validation_steps: The number of steps between validation checks for the HF Trainer arguments. pretrained_model: Pretrained model to use for fine-tuning. Uses default of TinyLlama.