Code

feature: lightweight probe defaults#1116

Code

Draft

leondz wants to merge 20 commits intoNVIDIA:mainNVIDIA/garak:mainfrom leondz:feature/lightweight_probe_defaultsleondz/garacc:feature/lightweight_probe_defaultsCopy head branch name to clipboard

+471-110

Commits

Commits on Feb 24, 2025

• stabilise and make explicit order of multiple Migrations in one fixer class

  leondz

  committed
  640d304

  Copy full SHA for 640d304

• update FigStep names

  leondz

  committed
  e7c7db5

  Copy full SHA for e7c7db5

• rename and fixers for snowball

  leondz

  committed
  a85c257

  Copy full SHA for a85c257

Commits on Feb 26, 2025

• add config entry for soft cap on how many prompts per probe

  leondz

  committed
  4370779

  Copy full SHA for 4370779

• rename promptinject probes & bind to soft probe prompt cap

  leondz

  committed
  169d481

  Copy full SHA for 169d481

• migrate past tense probe names

  leondz

  committed
  2e40865

  Copy full SHA for 2e40865

• resolve merge

  leondz

  committed
  829f97f

  Copy full SHA for 829f97f

• rename probes to have lightweight versions as defaults and extended/full versions with special names

  leondz

  committed
  ea5bed8

  Copy full SHA for ea5bed8

• shrink LatentInjectionFactSnippetEiffel to soft cap, w/ shuffle

  leondz

  committed
  79105cb

  Copy full SHA for 79105cb

• rename FalseAssertion, Glitch, use soft cap

  leondz

  committed
  3aa6677

  Copy full SHA for 3aa6677

• fix rename

  leondz

  committed
  3b3e786

  Copy full SHA for 3b3e786

• get order of operations right: set max_prompts after _config is available

  leondz

  committed
  88411ab

  Copy full SHA for 88411ab

• lightweight defaults for latent injection probes

  leondz

  committed
  3f03bc2

  Copy full SHA for 3f03bc2

• use random shuffle + prune for lightweight slur continuation

  leondz

  committed
  0375e52

  Copy full SHA for 0375e52

• move to using shuffling & prompt cap to produce lightweight probes

  leondz

  committed
  02d202a

  Copy full SHA for 02d202a

• access config_root not _config

  leondz

  committed
  7310811

  Copy full SHA for 7310811

Commits on Feb 27, 2025

• fixer class sorting should.. work

  leondz

  committed
  e28f8c0

  Copy full SHA for e28f8c0

• update test cases to fit current state of class names

  leondz

  committed
  f45938b

  Copy full SHA for f45938b

• constrain class replacement to final position in plugin name

  leondz

  committed
  0163a4c

  Copy full SHA for 0163a4c

Commits on Feb 28, 2025

• place migrations involving ordered ops into single classes. much tidier

  leondz

  committed
  f5d168a

  Copy full SHA for f5d168a