laya-multilingual-CoreAI / validation.json
ahmedandaloes's picture
Add laya-multilingual Apple Core AI conversion (macOS 27, measured)
b347e24 verified
Raw History Blame Contribute Delete
2.39 kB
{
"model": "laya",
"bundle": "bundles/laya-multilingual-CoreAI",
"reference": "laya package eager PyTorch forward on CPU (same build_sequence tokens)",
"cases": [
{
"case": 1,
"kind": "choice",
"options": 3,
"state_tokens": 56,
"reference_choice": "billing",
"coreai_choice": "billing",
"selected_answer_matches": true,
"max_probability_difference": 3.73371333850514e-11,
"max_abs_logit_difference": 4.482269287109375e-05
},
{
"case": 2,
"kind": "choice",
"options": 3,
"state_tokens": 51,
"reference_choice": "technical",
"coreai_choice": "technical",
"selected_answer_matches": true,
"max_probability_difference": 1.3395307751906671e-08,
"max_abs_logit_difference": 4.76837158203125e-07
},
{
"case": 3,
"kind": "noul",
"options": 2,
"state_tokens": 48,
"reference_choice": "false",
"coreai_choice": "false",
"selected_answer_matches": true,
"max_probability_difference": 1.3820633637529056e-08,
"max_abs_logit_difference": 9.5367431640625e-07
},
{
"case": 4,
"kind": "noul",
"options": 2,
"state_tokens": 47,
"reference_choice": "false",
"coreai_choice": "false",
"selected_answer_matches": true,
"max_probability_difference": 3.266735856266578e-07,
"max_abs_logit_difference": 7.152557373046875e-06
},
{
"case": 5,
"kind": "score",
"options": 3,
"state_tokens": 50,
"reference_choice": "blocking",
"coreai_choice": "blocking",
"selected_answer_matches": true,
"max_probability_difference": 4.5500904821693666e-08,
"max_abs_logit_difference": 3.5762786865234375e-06
},
{
"case": 6,
"kind": "score",
"options": 3,
"state_tokens": 45,
"reference_choice": "blocking",
"coreai_choice": "blocking",
"selected_answer_matches": true,
"max_probability_difference": 2.3378533586804195e-07,
"max_abs_logit_difference": 1.430511474609375e-06
}
],
"summary": {
"cases": 6,
"all_selected_answers_match": true,
"max_probability_difference": 3.266735856266578e-07,
"note": "Laya's own decision cases on one checkpoint. Not a benchmark of accuracy or calibration; it measures conversion drift only."
}
}