{ "ckpt": "ckpts/V_std_best.pt", "limit": 1000, "results": { "hellaswag": 0.398, "piqa": 0.69, "arc_easy": 0.5175, "arc_challenge": 0.3043, "winogrande": 0.541, "boolq": 0.491, "sciq": 0.62, "lambada": 0.334, "mean": 0.487 } }