downstep / calibration.json
NagaYu's picture
Downstep 35M: calibrated pitch accent for dictionary-unlisted Japanese compounds
04759fa verified
Raw History Blame Contribute Delete
644 Bytes
{
"fitted_on": "benchmarks/_raw_dev_unlisted.json",
"n_dev_items": 5441,
"temperature": 1.4978117508891504,
"ece_before": 0.048161102888937414,
"ece_after": 0.011165273365421908,
"thresholds": {
"max_accuracy": {
"threshold": 0.318732,
"dev_combined_accuracy": 0.8213563683146481
},
"zero_regression": {
"threshold": 0.984069,
"dev_coverage": 0.07755927219261165
}
},
"default_threshold": 0.318732,
"note": "Fitted on dev only. scripts/evaluate.py reports the real test-set regression count at these thresholds; a dev-set zero-regression point is an empirical guarantee, not a proof."
}