| { | |
| "digits": 12, | |
| "inference_time_on_baseline": { | |
| "confidence": [ | |
| 0.766875, | |
| 0.764375, | |
| 0.76375, | |
| 0.763125, | |
| 0.7625 | |
| ], | |
| "entropy": [ | |
| 0.766875, | |
| 0.764375, | |
| 0.76375, | |
| 0.763125, | |
| 0.7625 | |
| ], | |
| "entropy-gated": [ | |
| 0.766875, | |
| 0.764375, | |
| 0.76375, | |
| 0.761875, | |
| 0.7625 | |
| ], | |
| "margin": [ | |
| 0.766875, | |
| 0.764375, | |
| 0.76375, | |
| 0.763125, | |
| 0.761875 | |
| ], | |
| "random": [ | |
| 0.766875, | |
| 0.768125, | |
| 0.77125, | |
| 0.764375, | |
| 0.77 | |
| ] | |
| }, | |
| "ours": { | |
| "confidence": [ | |
| 1.0, | |
| 1.0, | |
| 1.0, | |
| 0.999375, | |
| 0.999375 | |
| ] | |
| }, | |
| "significance": { | |
| "confidence": { | |
| "diff": 0.233125, | |
| "ci95": [ | |
| 0.226875, | |
| 0.23875 | |
| ], | |
| "p_value": 0.0, | |
| "n": 400, | |
| "their_best_budget": 1 | |
| }, | |
| "entropy": { | |
| "diff": 0.233125, | |
| "ci95": [ | |
| 0.226875, | |
| 0.23875 | |
| ], | |
| "p_value": 0.0, | |
| "n": 400, | |
| "their_best_budget": 1 | |
| }, | |
| "entropy-gated": { | |
| "diff": 0.233125, | |
| "ci95": [ | |
| 0.226875, | |
| 0.23875 | |
| ], | |
| "p_value": 0.0, | |
| "n": 400, | |
| "their_best_budget": 1 | |
| }, | |
| "margin": { | |
| "diff": 0.233125, | |
| "ci95": [ | |
| 0.226875, | |
| 0.23875 | |
| ], | |
| "p_value": 0.0, | |
| "n": 400, | |
| "their_best_budget": 1 | |
| }, | |
| "random": { | |
| "diff": 0.22875, | |
| "ci95": [ | |
| 0.221875, | |
| 0.235625 | |
| ], | |
| "p_value": 0.0, | |
| "n": 400, | |
| "their_best_budget": 4 | |
| } | |
| }, | |
| "summary": { | |
| "best_inference_1pass": 0.766875, | |
| "best_inference_any_budget": 0.77125, | |
| "ours_1pass": 1.0 | |
| } | |
| } |