---

name: converters-datasets description: "Use PyRIT converters, converter stacks, message normalization, and seed datasets safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

PyRIT converters and datasets

Use this sub-skill when an agent needs to transform prompt/message values, compose converter stacks, normalize PyRIT messages for targets, or create/load seed datasets for PyRIT 1.1.0.dev0.

Start here

1. For direct offline transformation, choose a converter and call convert_async() or convert_tokens_async() using the contracts in references/converters-and-datasets.md.
2. For stacks passed into attacks, build ConverterConfiguration objects here, then route attack execution to ../attacks-scenarios/SKILL.md.
3. For target sends, target capability checks, and scorer configuration, route to ../targets-scorers/SKILL.md. This sub-skill only prepares messages, converters, and datasets.
4. For session initialization, memory setup, registry/default configuration, and persistent seed storage, route to ../setup-memory-core/SKILL.md.
5. For CLI dataset listing or backend scanner commands, route to ../cli-backend-scanner/SKILL.md.

Runtime references and helper

• Converter and dataset workflows covers converter modalities, offline versus service-backed converters, converter stacks, message normalizers, and dataset provider APIs.
• Data formats and YAML schemas covers PromptDataType, Message/MessagePiece, SeedPrompt, SeedDataset, seed YAML shapes, template parameters, and bundled response JSON schemas.
• Troubleshooting covers modality mismatches, optional media dependencies, LLM-backed converter target requirements, regex errors, YAML validation, remote dataset/cache failures, and missing template parameters.
• scripts/converter_dataset_smoke.py is a no-secret/no-network smoke helper that imports installed PyRIT, inspects core signatures, exercises Base64Converter, SearchReplaceConverter, SeedPrompt, and a tiny local SeedDataset, and prints a JSON summary.

Safety and boundary reminders

• Prefer offline converters for smoke tests and deterministic debugging. Base64Converter and SearchReplaceConverter are safe offline examples.
• LLM-backed converters call a converter_target and therefore may use credentials, rate limits, network, and target-specific chat capabilities. Do not instantiate them for a no-secret smoke unless a target has already been configured and authorized.
• Remote dataset providers can download from public URLs, HuggingFace, or ZIP archives and may populate PyRIT's data cache. Use local YAML or explicit small dataset_names for bounded work.
• NoOpConverter is not exported from pyrit.converter in this checkout; use an empty converter list when no conversion is needed.