configs/final_eval_set.json

{
  "name": "supplymind_final_eval_set_v1",
  "description": "Small cross-tier held-out benchmark for final SupplyMind model comparisons.",
  "cases": [
    {
      "tier": "easy",
      "task_id": "v2_train_easy",
      "seed": 131
    },
    {
      "tier": "medium",
      "task_id": "v2_train_medium",
      "seed": 211
    },
    {
      "tier": "hard",
      "task_id": "v2_train_hard",
      "seed": 307
    }
  ],
  "compare_setups": [
    "base_center_base_warehouse",
    "sft_center_sft_warehouse",
    "grpo_center_best_warehouse",
    "best_center_best_warehouse"
  ],
  "metrics": [
    "global_score",
    "center_role_score",
    "warehouse_role_score",
    "raw_reward",
    "invalid_payloads",
    "invalid_actions"
  ]
}