metrics: quantization parity benchmark (unified_v3 test split)
Browse files- metrics/quant_bench.json +94 -0
metrics/quant_bench.json
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"dataset": "unified_v3 test split",
|
| 3 |
+
"max_length": 128,
|
| 4 |
+
"n_rows": 2914,
|
| 5 |
+
"reference_fp32": {
|
| 6 |
+
"loss": 0.053679560930409015,
|
| 7 |
+
"n": 2914,
|
| 8 |
+
"accuracy": 0.9543582704186685,
|
| 9 |
+
"f1": 0.9645043898167938,
|
| 10 |
+
"precision": 0.9655916791519741,
|
| 11 |
+
"recall": 0.963423773743917,
|
| 12 |
+
"confusion_matrix": [
|
| 13 |
+
[
|
| 14 |
+
[
|
| 15 |
+
2233,
|
| 16 |
+
15
|
| 17 |
+
],
|
| 18 |
+
[
|
| 19 |
+
19,
|
| 20 |
+
647
|
| 21 |
+
]
|
| 22 |
+
],
|
| 23 |
+
[
|
| 24 |
+
[
|
| 25 |
+
2041,
|
| 26 |
+
27
|
| 27 |
+
],
|
| 28 |
+
[
|
| 29 |
+
26,
|
| 30 |
+
820
|
| 31 |
+
]
|
| 32 |
+
],
|
| 33 |
+
[
|
| 34 |
+
[
|
| 35 |
+
2353,
|
| 36 |
+
26
|
| 37 |
+
],
|
| 38 |
+
[
|
| 39 |
+
27,
|
| 40 |
+
508
|
| 41 |
+
]
|
| 42 |
+
]
|
| 43 |
+
],
|
| 44 |
+
"per_class": {
|
| 45 |
+
"source:sensitive": {
|
| 46 |
+
"precision": 0.9773413897280967,
|
| 47 |
+
"recall": 0.9714714714714715,
|
| 48 |
+
"f1": 0.9743975903614458,
|
| 49 |
+
"support": 666
|
| 50 |
+
},
|
| 51 |
+
"source:untrusted": {
|
| 52 |
+
"precision": 0.9681227863046045,
|
| 53 |
+
"recall": 0.9692671394799054,
|
| 54 |
+
"f1": 0.9686946249261665,
|
| 55 |
+
"support": 846
|
| 56 |
+
},
|
| 57 |
+
"sink:external": {
|
| 58 |
+
"precision": 0.951310861423221,
|
| 59 |
+
"recall": 0.9495327102803738,
|
| 60 |
+
"f1": 0.950420954162769,
|
| 61 |
+
"support": 535
|
| 62 |
+
}
|
| 63 |
+
}
|
| 64 |
+
},
|
| 65 |
+
"variants": {
|
| 66 |
+
"fp16": {
|
| 67 |
+
"n": 2914,
|
| 68 |
+
"source:sensitive": 0.9743,
|
| 69 |
+
"source:untrusted": 0.9693,
|
| 70 |
+
"sink:external": 0.9504,
|
| 71 |
+
"f1_macro": 0.9647,
|
| 72 |
+
"latency_s_total": 58.6,
|
| 73 |
+
"ms_per_text": 20.12
|
| 74 |
+
},
|
| 75 |
+
"int8": {
|
| 76 |
+
"n": 2914,
|
| 77 |
+
"source:sensitive": 0.9766,
|
| 78 |
+
"source:untrusted": 0.9697,
|
| 79 |
+
"sink:external": 0.953,
|
| 80 |
+
"f1_macro": 0.9664,
|
| 81 |
+
"latency_s_total": 55.1,
|
| 82 |
+
"ms_per_text": 18.9
|
| 83 |
+
},
|
| 84 |
+
"int8_int4_embeddings": {
|
| 85 |
+
"n": 2914,
|
| 86 |
+
"source:sensitive": 0.9789,
|
| 87 |
+
"source:untrusted": 0.9697,
|
| 88 |
+
"sink:external": 0.95,
|
| 89 |
+
"f1_macro": 0.9662,
|
| 90 |
+
"latency_s_total": 55.1,
|
| 91 |
+
"ms_per_text": 18.92
|
| 92 |
+
}
|
| 93 |
+
}
|
| 94 |
+
}
|