| { | |
| "dev_baseline": { | |
| "event": "eval", | |
| "split": "dev", | |
| "epoch": 0, | |
| "step": 0, | |
| "loss": 0.6314, | |
| "tool_output_filter": { | |
| "n": 21977, | |
| "pos": 20282, | |
| "neg": 1695, | |
| "auc": 0.7516, | |
| "tpr@0.5": 0.7605, | |
| "fpr@0.5": 0.3687, | |
| "tpr@0.9": 0.3445, | |
| "fpr@0.9": 0.1493, | |
| "tpr@0.99": 0.0, | |
| "fpr@0.99": 0.0 | |
| }, | |
| "action_authorization_guard": { | |
| "n": 0, | |
| "pos": 0, | |
| "neg": 0, | |
| "auc": 0.0, | |
| "tpr@0.35": 0.0, | |
| "fpr@0.35": 0.0, | |
| "tpr@0.5": 0.0, | |
| "fpr@0.5": 0.0, | |
| "tpr@0.9": 0.0, | |
| "fpr@0.9": 0.0 | |
| }, | |
| "elapsed_min": 8.46, | |
| "ts": 1793858877.13 | |
| }, | |
| "dev_best": { | |
| "event": "eval", | |
| "split": "dev_best", | |
| "epoch": 3, | |
| "step": 3047, | |
| "loss": 0.1934, | |
| "tool_output_filter": { | |
| "n": 21977, | |
| "pos": 20282, | |
| "neg": 1695, | |
| "auc": 0.984, | |
| "tpr@0.5": 0.9391, | |
| "fpr@0.5": 0.0501, | |
| "tpr@0.9": 0.8745, | |
| "fpr@0.9": 0.0206, | |
| "tpr@0.99": 0.6972, | |
| "fpr@0.99": 0.0029 | |
| }, | |
| "action_authorization_guard": { | |
| "n": 0, | |
| "pos": 0, | |
| "neg": 0, | |
| "auc": 0.0, | |
| "tpr@0.35": 0.0, | |
| "fpr@0.35": 0.0, | |
| "tpr@0.5": 0.0, | |
| "fpr@0.5": 0.0, | |
| "tpr@0.9": 0.0, | |
| "fpr@0.9": 0.0 | |
| } | |
| }, | |
| "test_banking_slack": { | |
| "event": "eval", | |
| "split": "test_banking_slack", | |
| "epoch": 3, | |
| "step": 3047, | |
| "loss": 0.1888, | |
| "tool_output_filter": { | |
| "n": 21965, | |
| "pos": 20271, | |
| "neg": 1694, | |
| "auc": 0.9833, | |
| "tpr@0.5": 0.9402, | |
| "fpr@0.5": 0.0543, | |
| "tpr@0.9": 0.8821, | |
| "fpr@0.9": 0.0201, | |
| "tpr@0.99": 0.7055, | |
| "fpr@0.99": 0.0041 | |
| }, | |
| "action_authorization_guard": { | |
| "n": 0, | |
| "pos": 0, | |
| "neg": 0, | |
| "auc": 0.0, | |
| "tpr@0.35": 0.0, | |
| "fpr@0.35": 0.0, | |
| "tpr@0.5": 0.0, | |
| "fpr@0.5": 0.0, | |
| "tpr@0.9": 0.0, | |
| "fpr@0.9": 0.0 | |
| }, | |
| "elapsed_min": 75.4, | |
| "checkpoint": { | |
| "path": "/root/work/checkpoints/clef-flash-LiteRT-ft-v04", | |
| "epoch": 3, | |
| "step": 3047 | |
| }, | |
| "ts": 1793862893.37 | |
| } | |
| } |