cap-sweep-eval-data

Anonymous supplementary release for a double-blind workshop submission. This dataset contains the raw per-task evaluation results and task definitions behind the paper's per-turn tool-call-cap sweep on the Opaque Knapsack, Navigation, and Rule Diagnosis agentic tasks.

Structure

Mirrors the experiments/cap_sweep/ layout expected by the (forthcoming anonymized) analysis code, so this can be dropped directly into a checkout at that path:

experiments/cap_sweep/
├── full2x2/        # 2x2 train-regime x runtime x {cap25, cap80}, knapsack, n=25/cell
├── dense/           # dense cap grid {10,20,25,30,40,50,80,100}, knapsack
├── seeds/           # 2x2 full grid replicated at 2 additional training seeds
├── second_family/    # navigation + rule_diagnosis generalization checks
├── tmax/             # turn-horizon control (T in {40,80,160}), rules out
│                      # "just too few turns" as the explanation for the
│                      # binding-cap collapse
├── tasks_knap_med/tasks/knapsack/       # knapsack task definitions
├── tasks_nav_med/tasks/navigation/      # navigation task definitions
└── tasks_rule_off/tasks/rule_diagnosis/ # rule-diagnosis task definitions

Each result cell contains per-task {family}-{family}-<id>.json (final outcome + metrics) and {family}-{family}-<id>.trace.json (full event-level tool-call trace) files, plus a summary_*.json aggregate and the exact YAML config used to produce it.

Provenance

Released anonymously alongside a NeurIPS workshop submission for reproducibility review. Non-anonymous release (paper citation, full code) will follow after the review process concludes.