---

name: pplm_recovery_evaluation description: Use when validating a reduced or full PPLM recovery experiment with numeric metrics and mechanism checks.

PPLM Recovery Evaluation

Use this skill to turn PPLM generation traces into recovery metrics.

Inputs

• Prompt/attribute items.
• Controlled and uncontrolled outputs.
• Trace flags for attribute model, perturbation, KL, reranking, and frozen base model.
• Paper target metadata.

Outputs

• Topic or sentiment hit rate.
• Diversity metrics.
• Mechanism checks for recovery_result.json.

Workflow

1. Canonicalize target words and generated tokens.
2. Compute attribute hit rate.
3. Aggregate diversity.
4. Require mechanism checks for proxy acceptance.
5. Emit validator-compatible JSON.

Validation

Run python tests/test_recovery_evaluation.py.

Limitations

Reduced proxy implementations must declare that they are not full GPT-2 345M reproduction.