Configurable TAP probe, refactor judge resources to shared red team

[jmartin-tech]

• resources for red team type activities as a shared lib
  • Model-as-a-judge logic can apply to many techniques
• refactor TAP probe to use Configurable
  • updated probe to accept generic model config
  • set default to cpu for hf model in tap generation
  • update util function to allow any valid generator
  • warn when generator is not known supported config
  • remove unused max_token params

Verification

List the steps needed to make sure this thing works

config.file.yaml:

plugins:
  generators:
    huggingface:
      hf_args:
        device: cpu
  probes:
    tap:
      attack_model_config:
        max_tokens: 500
        hf_args:
          device: cpu
      evaluator_model_type: nim
      evaluator_model_name: meta/llama3-8b-instruct
      evaluator_model_config:
        uri: http://0.0.0.0:8000/v1
        api_key: fake

• garak -m huggingface -n meta-llama/Llama-2-7b-chat-hf --config config.file.yaml --probes tap.TAP
• Run the tests and ensure they pass python -m pytest tests/

[leondz]

Thanks!! Before reviewing - is the scope a full LLMaaJ implementation or a first phase?

[jmartin-tech]

This is just foundation that will aid code reuse in the detector implementation, the PR enables the TAP probe to accept model configuration in it's existing usage of LLMaaJ.

[leondz]

minor comments re: tests, constants, deps; generally lgtm

[leondz]

lgtm