---

name: ranking_nce_estimator description: Optimize finite ranking NCE objectives and recover conditional distributions when partition functions vary by input.

Ranking NCE Estimator

Use this skill when a recovery needs the ranking objective from conditional NCE. It computes the paper's candidate-set objective and can optimize a small finite protocol without relying on an original implementation repository.

Do not use this skill for binary data-vs-noise NCE; use binary_nce_estimator for that objective.

Inputs

• A ConditionalNCEProtocol from conditional_nce_protocol.
• A score function or the built-in Section 4.3 two-parameter score.
• Candidate events from exact enumeration or sampled data.
• Bounded optimizer settings.

Outputs

• Ranking objective value.
• Estimated conditionals after score normalization over labels.
• Optimizer trace with parameters, loss values, and ratio diagnostics.
• Candidate posterior normalization checks.

Workflow

1. Compute adjusted scores bar_s=s-log p_N.
2. For each event, form a (K+1)-way softmax over the positive candidate and all negative candidates.
3. Accumulate the expected log probability of the positive candidate.
4. Normalize scores over the complete label set to produce p_hat(y|x).
5. In reduced recovery, optimize the Section 4.3 population objective and verify the x1 ratio approaches 1/3.

Validation

Run:

python /share/project/yuyang/workspace/Paper2Skills/Distiller/skills/module-to-skill/scripts/validate_skill_tree.py . --run-tests

The tests check objective finiteness, softmax normalization, and recovery of the Section 4.3 ratio.

Limitations

The included optimizer is designed for deterministic small-support recovery. Large neural language-model training should use a real ML stack and treat this skill as a semantic reference for the objective.