{
  "experiment": "A_B_downstream_crosslingual",
  "timestamp": "2026-04-13 09:59:30 UTC",
  "base": {
    "sentiment": {
      "he": {
        "logprob_acc": 0.53,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "ar": {
        "logprob_acc": 0.45,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "fa": {
        "logprob_acc": 0.605,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "en": {
        "logprob_acc": 0.515,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      }
    },
    "belebele": {
      "he": {
        "accuracy": 0.26,
        "n": 100
      },
      "ar": {
        "accuracy": 0.22,
        "n": 100
      },
      "fa": {
        "accuracy": 0.19,
        "n": 100
      },
      "en": {
        "accuracy": 0.19,
        "n": 100
      }
    }
  },
  "en_sft": {
    "sentiment": {
      "he": {
        "logprob_acc": 0.515,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "ar": {
        "logprob_acc": 0.465,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "fa": {
        "logprob_acc": 0.585,
        "logprob_n": 200,
        "gen_acc": 0.0,
        "gen_n": 50
      },
      "en": {
        "logprob_acc": 0.52,
        "logprob_n": 200,
        "gen_acc": 0.02,
        "gen_n": 50
      }
    },
    "belebele": {
      "he": {
        "accuracy": 0.26,
        "n": 100
      },
      "ar": {
        "accuracy": 0.23,
        "n": 100
      },
      "fa": {
        "accuracy": 0.2,
        "n": 100
      },
      "en": {
        "accuracy": 0.2,
        "n": 100
      }
    }
  },
  "multilingual_sft": {
    "sentiment": {
      "he": {
        "logprob_acc": 0.845,
        "logprob_n": 200,
        "gen_acc": 0.82,
        "gen_n": 50
      },
      "ar": {
        "logprob_acc": 0.605,
        "logprob_n": 200,
        "gen_acc": 0.64,
        "gen_n": 50
      },
      "fa": {
        "logprob_acc": 0.785,
        "logprob_n": 200,
        "gen_acc": 0.74,
        "gen_n": 50
      },
      "en": {
        "logprob_acc": 0.73,
        "logprob_n": 200,
        "gen_acc": 0.64,
        "gen_n": 50
      }
    },
    "belebele": {
      "he": {
        "accuracy": 0.28,
        "n": 100
      },
      "ar": {
        "accuracy": 0.24,
        "n": 100
      },
      "fa": {
        "accuracy": 0.25,
        "n": 100
      },
      "en": {
        "accuracy": 0.23,
        "n": 100
      }
    }
  },
  "generation_samples": [
    {
      "prompt": "Write a short sentence in Hebrew about the weather.",
      "response": "-הייתי עייף."
    },
    {
      "prompt": "اكتب جملة قصيرة عن الطقس بالعربية.",
      "response": "- I'm sorry."
    },
    {
      "prompt": "یک جمله کوتاه درباره آب و هوا به فارسی بنویسید.",
      "response": "- What is the weather like in the summer?"
    },
    {
      "prompt": "What is the capital of Israel?",
      "response": "מה זה?"
    }
  ]
}