analysis_outputs/task4_global/task4_raw_results.json

{
  "4": {
    "bertscore_f1": 0.6790409088134766,
    "semantic_sim": 0.5334163904190063,
    "bleu": 0.0003333366738921545,
    "speed_per_sample": 0.12499853520479519
  },
  "8": {
    "bertscore_f1": 0.6395084261894226,
    "semantic_sim": 0.48110902309417725,
    "bleu": 1.7409952895149752e-156,
    "speed_per_sample": 0.2653733541647671
  },
  "16": {
    "bertscore_f1": 0.6821702718734741,
    "semantic_sim": 0.5372508764266968,
    "bleu": 0.00047544921470022843,
    "speed_per_sample": 0.53346393458487
  },
  "32": {
    "bertscore_f1": 0.4939882755279541,
    "semantic_sim": 0.19846703112125397,
    "bleu": 8.528327361662488e-233,
    "speed_per_sample": 1.0967584718746366
  }
}