Imported from yzj623/pplab (
ResearchAgent-Skill/skill_library/paperbench/sequential-neural-score-estimation/skill/variational_methods_sbi/sir_posterior_refinement/SKILL.md). Install upstream withnpx skills add yzj623/pplab --skill sir_posterior_refinement. Copyright stays with the author.
sir_posterior_refinement
Use this skill when a recovery or implementation needs the paper mechanism from Variational Methods for Simulation-Based Inference without reading the original repository. Do not use it as a full replacement for large-scale sbi or sbibm benchmark training when those runtimes are available.
Inputs
- Numeric simulator, posterior, likelihood, or validity data matching the module contract.
- A bounded runtime where deterministic scripts can be executed.
- Optional random seed for reproducible proxy checks.
Outputs
- JSON-serializable diagnostics from the module script.
- Numeric values that can be cross-checked by a recovery harness.
- Clear flags when the behavior is reduced or proxy-only.
Workflow
- Read the caller-provided data and keep source boundaries explicit.
- Run the deterministic script in
scripts/for the relevant SNVI component. - Check numeric invariants such as normalized weights, data accumulation, posterior movement, or validity correction.
- Return diagnostics and preserve logs for downstream recovery validation.
Validation
Run python scripts/ only through the documented tests, or run python -m pytest tests when pytest is available. The Distiller validator command is python <distiller>/module-to-skill/scripts/validate_skill_tree.py <skill_dir> --run-tests.
Limitations
This skill captures reusable SNVI mechanisms. It does not claim to reproduce the paper's full C2ST benchmark curves or pyloric-network experiment unless paired with the required packages, simulators, and compute budget.