auto-200m-2-int8 / eval /ptq_sweep.json
ProCreations's picture
Release auto-200m-2-int8 (150 MB, 2890/3000 on Approve-or-Deny)
2501a22 verified
Raw History Blame Contribute Delete
1.22 kB
{
"results": [
{
"name": "ptq-int8",
"bits": 8,
"group_size": null,
"nll": 0.07195428063703702,
"accuracy": 0.9803169734151329,
"false_approve_count": 70,
"false_deny_count": 84,
"auroc": 0.9973858492403456
},
{
"name": "ptq-int4-g32",
"bits": 4,
"group_size": 32,
"nll": 0.07148108854293608,
"accuracy": 0.9808282208588958,
"false_approve_count": 74,
"false_deny_count": 76,
"auroc": 0.997330817927398
},
{
"name": "ptq-int4-g64",
"bits": 4,
"group_size": 64,
"nll": 0.07348664240835465,
"accuracy": 0.9805725971370143,
"false_approve_count": 72,
"false_deny_count": 80,
"auroc": 0.9971573581832608
},
{
"name": "ptq-int4-g128",
"bits": 4,
"group_size": 128,
"nll": 0.07262665653736888,
"accuracy": 0.9792944785276073,
"false_approve_count": 76,
"false_deny_count": 86,
"auroc": 0.9970640271679527
}
],
"int4_group_size": 128,
"rule": "G chosen from {32, 64, 128} by post-training (MSE-clipped) quantization NLL on the validation selection split; the largest G within 2% of the best NLL."
}