Stdlib › Safety
mellea.stdlib.safety.guardian
Module: mellea.stdlib.safety.guardian
Risk checking with Guardian models.
Classes
class mellea.stdlib.safety.guardian.GuardianRisk()
Risk definitions as defined in https://github.com/ibm-granite/granite-guardian/blob/main/cookbooks/granite-guardian-3.2/quick_start_vllm.ipynb .
class mellea.stdlib.safety.guardian.GuardianCheck(risk: str = GuardianRisk.HARM, model_version: str = 'ibm-granite/granite-guardian-3.2-3b-a800m', device: str | None = None)
A Check for Risks based on local huggingface backend.
Constructor
Initializes a GuardianModel instance with the specified risk definition, model version, and device.Arguments
risk: The risk definition to check for, defaults to GuardianRisk.HARM.model_version: The version of the model, defaults to “ibm-granite/granite-guardian-3.2-3b-a800m”.device: The computational device to use (“cuda” for GPU, “mps” for Apple Silicon, or “cpu”), defaults to None. If not specified, the best available device will be automatically selected.
Methods
mellea.stdlib.safety.guardian.GuardianCheck._parse_output(output, input_len, tokenizer)
Parse the output of a guardian model and determine whether if the risk is present or not.
Arguments
output: The model’s output containing sequences from which predictions are made.input_len: The length of the original input sequence used for alignment with the model’s output.tokenizer: The tokenizer associated with the model, used to decode the tokens back into text.
mellea.stdlib.safety.guardian.GuardianCheck._guardian_validate(ctx: Context)
Validates the last turn of a conversation context using wrt given risk.
Code is adopted from https://huggingface.co/ibm-granite/granite-guardian-3.2-3b-a800m#quickstart-example
Arguments
ctx:Context: The context object containing the last turn of the conversation.
Returns
bool: True if there is no identified risk, False otherwise.