Installs into .claude/skills of the current project.
Are you the author of Pplm Latent Perturbation?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/vectorspacelab-pplm-latent-perturbation)
---
name: 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
1. Copy the base logits/state and keep it immutable.
2. Initialize a perturbation at zero.
3. Iteratively estimate the gradient of attribute loss plus KL-to-base penalty.
4. Update only the perturbation.
5. 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.
