1{2 "system": "p4v2_gr_s3",3 "split": "test_depth",4 "checkpoint": "experiments/runs/p4v2_gr_seed03_20261030001749/checkpoint",5 "checkpoint_sha256": "29c37958b4eefbbcba2b5fb04b3f1ee570c53678afafc28776ec659d5cbace17",6 "config_sha256": "a9cd9e1dabd9c0a5207678b323f88b0854bc936f59ffc8a0eae25664f80c3d09",7 "code_commit": "unknown",8 "device": "mps",9 "precision": "fp32",10 "evaluated_at": "2026-10-30T04:06:15-0400",11 "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",12 "exact_match": {13 "budget0": {14 "accuracy": 0.312,15 "n": 1000,16 "by_family": {17 "arith": 0.0,18 "bind": 0.02,19 "fsm": 0.425,20 "excl": 0.81,21 "sched": 0.30522 },23 "edits_per_token": 0.024 },25 "forced_b2": {26 "accuracy": 0.332,27 "n": 1000,28 "by_family": {29 "arith": 0.0,30 "bind": 0.045,31 "fsm": 0.47,32 "excl": 0.78,33 "sched": 0.36534 },35 "edits_per_token": 2.036 }37 },38 "seconds": 11.839}