brainstorming
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).
Explorer les Skills
Développement · SKILL.md
pplm_latent_perturbation
pplm_latent_perturbationUse when a PPLM-style controller must perturb latent/logit state at inference time with attribute and KL losses while freezing the base model.
Independent GitHub snapshot · Automated classification · Original documentation governs
Documentation de la Skill
GitHubname: pplm_latent_perturbation description: Use when a PPLM-style controller must perturb latent/logit state at inference time with attribute and KL losses while freezing the base model.
PPLM Latent Perturbation
Use this skill to implement the core PPLM inference-time update.
Inputs
- Base next-token logits or latent state.
- Attribute loss function.
- Step size, iteration count, KL scale.
Outputs
- Perturbed probability distribution.
- Trace with before/after loss, KL, and perturbation values.
Workflow
- Copy the base logits/state and keep it immutable.
- Initialize a perturbation at zero.
- Iteratively estimate the gradient of attribute loss plus KL-to-base penalty.
- Update only the perturbation.
- Return the final distribution and trace proving the target probability changed.
Validation
Run python tests/test_latent_perturbation.py.
Limitations
Reduced proxy implementations must declare that they are not full GPT-2 345M reproduction.