calibration.json

{
 "temperatures_per_question_type": {
  "choice": 1.0971,
  "score": 1.139,
  "noul": 1.162
 },
 "temperature_fit_items": {
  "choice": 2533,
  "score": 261,
  "noul": 2031
 },
 "note": "The temperatures are already applied by laya (rl_agent_config.json 'temperature') and baked into the GGUF files. They were fitted on held-out training items, never on the test sets.",
 "scam_threshold": {
  "question": "noul, 'is this message a scam'",
  "threshold": 0.4861,
  "chosen_on": "1188 held-out training messages, never the test",
  "held_out_acc_at_threshold": 0.979,
  "held_out_pools": {
   "held-out warnings about scams, short scams without a link and look-alike messages": {
    "n": 188,
    "acc_at_threshold": 0.984
   },
   "held-out hard scam and legitimate messages": {
    "n": 300,
    "acc_at_threshold": 0.9933
   },
   "held-out spam messages": {
    "n": 700,
    "acc_at_threshold": 0.9714
   }
  },
  "test_acc_at_0.5": 0.9195,
  "test_acc_at_threshold": 0.9228,
  "test_hard_acc_at_0.5": 0.8308,
  "test_hard_acc_at_threshold": 0.8308,
  "test_auc": 0.969
 },
 "ramzor_example_thresholds": {
  "green_below": 0.35,
  "red_from": 0.49,
  "zones_on_test": {
   "g": {
    "n": 207,
    "share": 0.6946308724832215,
    "scam": 13,
    "legit": 194
   },
   "y": {
    "n": 9,
    "share": 0.030201342281879196,
    "scam": 2,
    "legit": 7
   },
   "r": {
    "n": 82,
    "share": 0.2751677852348993,
    "scam": 73,
    "legit": 9
   }
  }
 }
}