artifacts/evaluation/model_comparison/model_comparison.json

{
  "Majority Baseline": {
    "average_accuracy": 0.42628087386822827,
    "average_macro_f1": 0.07927692465823664,
    "average_top3_accuracy": 0.7344469174752037,
    "exact_match_accuracy": 0.009529571925578581,
    "latency_ms": 0.0005647709000186296
  },
  "Frozen CLIP + Heads (V1)": {
    "average_accuracy": 0.8335026038852995,
    "average_macro_f1": 0.6568447666058456,
    "average_top3_accuracy": 0.9711087581303888,
    "exact_match_accuracy": 0.27938284677053393,
    "latency_ms": 6.226557110003341
  },
  "AutoCatalogAI V2": {
    "average_accuracy": 0.8747758065561727,
    "average_macro_f1": 0.6740687827687263,
    "average_top3_accuracy": 0.981502690321326,
    "exact_match_accuracy": 0.4046286492209953,
    "latency_ms": 7.4737231199719645
  }
}