Datasets:

garn-garn
/

smartgridbench-review-artifact

license: cc-by-4.0
pretty_name: SmartGridBench Review Artifact
language:
  - en
tags:
  - smart-grid
  - benchmark
  - tool-use
  - model-context-protocol
  - llm-evaluation
configs:
  - config_name: evidence_registry
    data_files:
      - split: train
        path: data/evidence_registry.csv
  - config_name: failure_taxonomy_current
    data_files:
      - split: train
        path: data/failure_taxonomy_current.csv
  - config_name: failure_taxonomy_current_auto_label_counts
    data_files:
      - split: train
        path: data/failure_taxonomy_current_auto_label_counts.csv
  - config_name: failure_taxonomy_current_failed_dim_counts
    data_files:
      - split: train
        path: data/failure_taxonomy_current_failed_dim_counts.csv
  - config_name: failure_taxonomy_current_manual_audit_counts
    data_files:
      - split: train
        path: data/failure_taxonomy_current_manual_audit_counts.csv
  - config_name: failure_evidence_table
    data_files:
      - split: train
        path: data/failure_evidence_table.csv
  - config_name: gcp_post175_70b_summary
    data_files:
      - split: train
        path: data/gcp_post175_70b_summary.csv
  - config_name: gcp_post175_core31_summary
    data_files:
      - split: train
        path: data/gcp_post175_core31_summary.csv
  - config_name: gcp_post175_final_summary
    data_files:
      - split: train
        path: data/gcp_post175_final_summary.csv
  - config_name: gcp_post175_mitigation_4tier_summary
    data_files:
      - split: train
        path: data/gcp_post175_mitigation_4tier_summary.csv
  - config_name: manual_judge_audit
    data_files:
      - split: train
        path: data/manual_judge_audit.csv
  - config_name: mitigation_before_after
    data_files:
      - split: train
        path: data/mitigation_before_after.csv
  - config_name: mitigation_run_inventory
    data_files:
      - split: train
        path: data/mitigation_run_inventory.csv
  - config_name: paper_grade_scenario_scores
    data_files:
      - split: train
        path: data/paper_grade_scenario_scores.csv
  - config_name: profiling_inventory
    data_files:
      - split: train
        path: data/profiling_inventory.csv
  - config_name: scenario_catalog
    data_files:
      - split: train
        path: data/scenario_catalog.csv

SmartGridBench Review Artifact

This anonymous reviewer artifact contains the paper-facing SmartGridBench scenario and evidence tables for a NeurIPS 2026 Evaluations & Datasets submission.

SmartGridBench evaluates tool-using language agents on Smart Grid transformer-maintenance workflows. The package includes 36 canonical scenario definitions (31 hand-authored plus 5 promoted generated scenarios), 5 negative schema fixtures, paper-grade evidence registries, summary tables, a compact judge-score table, a manual judge audit table, and a compact profiling/W&B inventory with private links redacted.

data/scenario_catalog.csv: tabular catalog of the 36 canonical scenario JSON files.
scenarios/*.json: canonical scenario definitions: 31 hand-authored plus 5 promoted generated scenarios.
negative_checks/*.json: intentionally invalid schema fixtures for validator checks.
data/evidence_registry.csv: machine-readable ledger for paper-grade, superseded, diagnostic, and invalid cohorts.
data/gcp_post175_final_summary.csv: paper-facing summary for 37 row groups and 2,420 judged trajectories.
data/gcp_post175_core31_summary.csv: 31-scenario 8B core matrix summary.
data/gcp_post175_mitigation_4tier_summary.csv: matched mitigation ladder summary.
data/gcp_post175_70b_summary.csv: hosted 70B scaling summary.
data/mitigation_before_after.csv: current post-PR175 mitigation before/after comparison export.
data/profiling_inventory.csv: compact profiling/W&B inventory; private run URLs and artifact names are redacted.
data/paper_grade_scenario_scores.csv: compact per-trial judge score table for included paper-grade rows.
data/manual_judge_audit.csv: 12-row manual audit sanity check of the judge.
PROVENANCE.md: source, inclusion, exclusion, and anonymization notes.
VALIDATION.md: local package validation and leak-check notes.
MANIFEST.csv: file-level SHA-256 manifest for this artifact package.
code/: anonymized executable source snapshot with runners, MCP servers, configs, tests, and public-safe data fixtures.

Code Package

An anonymized executable code package is included at code/. Reviewers can browse it at:

https://huggingface.co/datasets/garn-garn/smartgridbench-review-artifact/tree/main/code

The code package includes its own README.md, LICENSE.md, and CODE_PACKAGE_VALIDATION.md. Code is MIT-licensed; the surrounding data/artifact package remains CC BY 4.0. Full benchmark execution requires model credentials and, for Plan-Execute paths, a reviewer-local AssetOpsBench checkout supplied via AOB_PATH.

Croissant Metadata

This package includes an anonymous Croissant JSON-LD metadata file at:

https://huggingface.co/datasets/garn-garn/smartgridbench-review-artifact/resolve/main/croissant.json

The Croissant metadata is a machine-readable description of the reviewer artifact. It is not a replacement for the Data Card, provenance notes, or paper limitations.

License

Dataset documentation, scenario definitions, and derived evidence tables in this reviewer artifact are released under CC BY 4.0. Code in the associated benchmark repository is licensed separately under MIT.

Version

Package version: neurips2026-review-v1
Package refreshed UTC: 2026-05-07T10:35:14Z
Package source snapshot SHA: c9c2b6c6a2c85d97570f37a3d5c77d3d9ef136a8
Anonymous code snapshot SHA: c9c2b6c6a2c85d97570f37a3d5c77d3d9ef136a8
Artifact import SHA: c9c2b6c6a2c85d97570f37a3d5c77d3d9ef136a8
Evidence capture floor SHA: 1913c6e4703425f735d8cb8297cb890ba66bbeff

SmartGridBench Review Artifact

Contents

Code Package

Croissant Metadata

License

Version