geppo / runtime.json
grokked-it's picture
Add GEPPO mean-pooling pointer-head verification release
e6742a5 verified
Raw History Blame Contribute Delete
3.96 kB
{
"primary": "mean-pointer-53",
"device": "mps",
"pooling": "mean",
"load_seconds": 5.890662916994188,
"validation_probes": 12,
"warm_median_single_question_ms": 62.09599999419879,
"choice_order_and_id_max_probability_error": 1.1102230246251565e-16,
"peak_process_rss_gib": 0.3963775634765625,
"driver_gib": 6.04754638671875,
"demo": {
"model": "geppo-emma-mean-pointer-53",
"answers": {
"reparto": {
"type": "choice",
"probabilities": {
"pagamenti": 0.2539552140965876,
"consegne": 0.49414000431610916,
"account": 0.2519047815873033
},
"calibrated": true,
"choice": "consegne"
},
"urgenza": {
"type": "noul",
"probabilities": {
"no": 0.720944137922657,
"yes": 0.27905586207734306
},
"calibrated": true,
"noul": 0.27905586207734306
},
"tono": {
"type": "score",
"probabilities": {
"0": 0.060208649775094075,
"1": 0.029866586670363942,
"2": 0.9099247635545419
},
"calibrated": false,
"score": 1.8497161137794478,
"legend": {
"0": "Calmo",
"1": "Frustrato",
"2": "Molto arrabbiato"
}
}
},
"calibration": {
"choice": {
"temperature": 2.602446831677766,
"applied": true,
"fit_questions": 90,
"audit_questions": 63,
"fitted_temperature": 2.602446831677766,
"audit_raw_nll": 0.8472414408314666,
"audit_adjusted_nll": 0.711108307300383
},
"noul": {
"temperature": 1.350695033667327,
"applied": true,
"fit_questions": 59,
"audit_questions": 49,
"fitted_temperature": 1.350695033667327,
"audit_raw_nll": 0.5295909861306163,
"audit_adjusted_nll": 0.5029942074278655
},
"score": {
"temperature": 1.0,
"applied": false,
"fit_questions": 6,
"audit_questions": 14
}
},
"note": "Experimental probabilities; no reliability guarantee on new domains."
},
"dictionary_state_demo": {
"model": "geppo-emma-mean-pointer-53",
"answers": {
"reparto": {
"type": "choice",
"probabilities": {
"pagamenti": 0.1965398838467544,
"consegne": 0.40428071054636444,
"account": 0.3991794056068812
},
"calibrated": true,
"choice": "consegne"
},
"urgenza": {
"type": "noul",
"probabilities": {
"no": 0.6338319200080171,
"yes": 0.36616807999198286
},
"calibrated": true,
"noul": 0.36616807999198286
},
"tono": {
"type": "score",
"probabilities": {
"0": 0.012831145717048503,
"1": 0.10413001435937663,
"2": 0.8830388399235748
},
"calibrated": false,
"score": 1.8702076942065262,
"legend": {
"0": "Calmo",
"1": "Frustrato",
"2": "Molto arrabbiato"
}
}
},
"calibration": {
"choice": {
"temperature": 2.602446831677766,
"applied": true,
"fit_questions": 90,
"audit_questions": 63,
"fitted_temperature": 2.602446831677766,
"audit_raw_nll": 0.8472414408314666,
"audit_adjusted_nll": 0.711108307300383
},
"noul": {
"temperature": 1.350695033667327,
"applied": true,
"fit_questions": 59,
"audit_questions": 49,
"fitted_temperature": 1.350695033667327,
"audit_raw_nll": 0.5295909861306163,
"audit_adjusted_nll": 0.5029942074278655
},
"score": {
"temperature": 1.0,
"applied": false,
"fit_questions": 6,
"audit_questions": 14
}
},
"note": "Experimental probabilities; no reliability guarantee on new domains."
}
}