Generates a red-team dataset by invoking one or more plugins.
Generation is separate from running. The output is a regular Tribunal
dataset file that you execute with mix tribunal.eval or tribunal_eval/2,
the same as any hand-written eval suite.
Usage
mix tribunal.redteam.generate \
--plugins policy \
--purpose-file priv/purpose.txt \
--policy-file priv/policy.txt \
--count 5 \
--output test/evals/datasets/redteam.yamlOptions
--plugins- Required. Comma-separated plugin ids (e.g.,policy).--purpose- Inline purpose text. Mutually exclusive with--purpose-file.--purpose-file- Path to a file containing the purpose text.--policy- Inline policy text.--policy-file- Path to a file containing the policy text.--count- Cases per plugin (default: 5).--output- Path to write the dataset. Format inferred from extension (.yaml/.yml→ YAML,.json→ JSON). Defaults to stdout in YAML.--format- Override format detection (yamlorjson).--model- Attacker LLM model spec, passed through to the attacker.
Example
mix tribunal.redteam.generate \
--plugins policy \
--purpose "Shopping assistant for a cosmetics retailer." \
--policy-file priv/policy.txt \
--count 5 \
--output test/evals/datasets/policy_redteam.yaml