laya-CoreAI / validation.json
ahmedandaloes's picture
Add laya Apple Core AI conversion (macOS 27, measured)
d3b64b7 verified
Raw History Blame Contribute Delete
2.4 kB
{
"model": "laya",
"bundle": "/Users/andaloes/mlx-work/laya/bundles/laya-CoreAI",
"reference": "laya package eager PyTorch forward on CPU (same build_sequence tokens)",
"cases": [
{
"case": 1,
"kind": "choice",
"options": 3,
"state_tokens": 59,
"reference_choice": "billing",
"coreai_choice": "billing",
"selected_answer_matches": true,
"max_probability_difference": 2.149687983177273e-09,
"max_abs_logit_difference": 6.67572021484375e-06
},
{
"case": 2,
"kind": "choice",
"options": 3,
"state_tokens": 55,
"reference_choice": "technical",
"coreai_choice": "technical",
"selected_answer_matches": true,
"max_probability_difference": 3.2404249949991026e-08,
"max_abs_logit_difference": 4.291534423828125e-06
},
{
"case": 3,
"kind": "noul",
"options": 2,
"state_tokens": 48,
"reference_choice": "true",
"coreai_choice": "true",
"selected_answer_matches": true,
"max_probability_difference": 4.996124104150734e-07,
"max_abs_logit_difference": 6.318092346191406e-06
},
{
"case": 4,
"kind": "noul",
"options": 2,
"state_tokens": 47,
"reference_choice": "false",
"coreai_choice": "false",
"selected_answer_matches": true,
"max_probability_difference": 4.186273309600419e-08,
"max_abs_logit_difference": 5.245208740234375e-05
},
{
"case": 5,
"kind": "score",
"options": 3,
"state_tokens": 45,
"reference_choice": "blocking",
"coreai_choice": "blocking",
"selected_answer_matches": true,
"max_probability_difference": 3.617309453085582e-07,
"max_abs_logit_difference": 8.940696716308594e-06
},
{
"case": 6,
"kind": "score",
"options": 3,
"state_tokens": 42,
"reference_choice": "soon",
"coreai_choice": "soon",
"selected_answer_matches": true,
"max_probability_difference": 6.151039953494752e-07,
"max_abs_logit_difference": 5.9604644775390625e-06
}
],
"summary": {
"cases": 6,
"all_selected_answers_match": true,
"max_probability_difference": 6.151039953494752e-07,
"note": "Laya's own decision cases on one checkpoint. Not a benchmark of accuracy or calibration; it measures conversion drift only."
}
}