---

name: pplm_candidate_reranking description: Use when generated candidates must be ranked by attribute likelihood with Dist-n diversity filtering as in PPLM BCR.

PPLM Candidate Reranking

Use this skill for PPLM BR/BCR-style candidate selection.

Inputs

• Candidate token sequences.
• Attribute scores where larger is better.
• Minimum Dist-n diversity threshold.

Outputs

• Dist-1/2/3 metrics.
• Filtered ranked list.
• Selected candidate and reason.

Workflow

1. Compute Dist-n as unique n-grams divided by total n-grams.
2. Filter candidates below the requested Dist threshold.
3. Select the highest attribute score among surviving candidates.
4. If all candidates fail diversity, select the highest score and record fallback.

Validation

Run python tests/test_candidate_reranking.py.

Limitations

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