dataset_info.json

{
  "dataset_name": "DesignBench-Truss",
  "version": "1.0",
  "description": "DesignBench is a benchmark for evaluating LLM capabilities in iterative structural engineering design. The truss domain uses finite element analysis (FEA) via trussme as the reward oracle. Each problem asks a model to minimize truss mass while satisfying buckling and yielding factor-of-safety constraints.",
  "num_problems": 10,
  "num_examples": 5000,
  "splits": {
    "benchmark": {
      "num_problems": 10,
      "num_examples": 5000,
      "purpose": "HuggingFace public benchmark — NEVER used for training"
    }
  },
  "grammar_actions": [
    "SCALE_PARAM",
    "SCALE_MULTI_PARAM",
    "ADD_MEMBER"
  ],
  "metrics": {
    "feasibility_rate": "fraction of episodes reaching is_feasible=True",
    "mass_reduction": "(initial_mass - final_mass) / initial_mass",
    "grammar_validity_rate": "fraction of steps with parseable grammar actions",
    "avg_steps_to_feasibility": "mean episode length when feasible",
    "overall_score": "weighted composite of design_correctness(0.5) + structural_validity(0.25) + reasoning_quality(0.15) + grammar_compliance(0.10)"
  }
}