open-system-one / data.json
dylantom2012's picture
Correct Laya's latency: it was Apple MPS, not CPU; add like-for-like CPU numbers
741f781 verified
Raw History Blame Contribute Delete
7.09 kB
{
"datasets": [
"sst2",
"ag_news",
"emotion",
"banking77"
],
"zero": [
{
"name": "typesafe/jev (bare labels)",
"acc": {
"sst2": 0.898,
"ag_news": 0.8828,
"emotion": 0.586,
"banking77": 0.7836
},
"lat": {
"sst2": 381.7,
"ag_news": 382.1,
"emotion": 374.0,
"banking77": 392.1
},
"ram": null,
"cloud": true
},
{
"name": "typesafe/jev (enriched)",
"acc": {
"sst2": 0.9164,
"ag_news": 0.8864,
"emotion": 0.59,
"banking77": 0.7784
},
"lat": {
"sst2": 381.7,
"ag_news": 382.1,
"emotion": 374.0,
"banking77": 392.1
},
"ram": null,
"cloud": true
},
{
"name": "Cross-encoder 149M (bare)",
"acc": {
"sst2": 0.8548,
"ag_news": 0.8444,
"emotion": 0.762,
"banking77": 0.6884
},
"lat": {
"sst2": 50.0,
"ag_news": 311.2,
"emotion": 216.0,
"banking77": 615.8
},
"ram": null
},
{
"name": "Cross-encoder 149M (enriched)",
"acc": {
"sst2": 0.8448,
"ag_news": 0.8464,
"emotion": 0.7172,
"banking77": 0.5688
},
"lat": {
"sst2": 71.9,
"ag_news": 378.7,
"emotion": 269.5,
"banking77": 707.4
},
"ram": null
},
{
"name": "Bi-encoder + 8-float head (bare)",
"acc": {
"sst2": 0.8112,
"ag_news": 0.6632,
"emotion": 0.4364,
"banking77": 0.6324
},
"lat": {
"sst2": 15.685,
"ag_news": 15.685,
"emotion": 15.685,
"banking77": 15.685
},
"ram": 490
},
{
"name": "Bi-encoder + 8-float head (enriched)",
"acc": {
"sst2": 0.8284,
"ag_news": 0.8,
"emotion": 0.5084,
"banking77": 0.6576
},
"lat": {
"sst2": 15.685,
"ag_news": 15.685,
"emotion": 15.685,
"banking77": 15.685
},
"ram": 490
},
{
"name": "Qwen2.5-0.5B generative",
"acc": {
"sst2": 0.844,
"ag_news": 0.6556,
"emotion": 0.3412,
"banking77": 0.142
},
"lat": {
"sst2": 103.5,
"ag_news": 110.1,
"emotion": 109.9,
"banking77": 314.2
},
"ram": 2382
},
{
"name": "Laya 421M (open weights, MPS)",
"acc": {
"sst2": 0.8652,
"ag_news": 0.9392,
"emotion": 0.582,
"banking77": 0.5428
},
"lat": {
"sst2": 19.4,
"ag_news": 25.8,
"emotion": 21.8,
"banking77": 86.3
},
"ram": null,
"device": "mps",
"note": "laya.load() auto-selects cuda>mps>cpu and chose mps; every other row here is pure CPU, batch=1. CPU re-measure in `laya_cpu`."
}
],
"few": [
{
"name": "static + 8-float head (2k labels)",
"acc": {
"sst2": 0.7832,
"ag_news": 0.8772,
"emotion": 0.6728,
"banking77": 0.7936
},
"lat": {
"sst2": 0.1067,
"ag_news": 0.1067,
"emotion": 0.1067,
"banking77": 0.1067
},
"ram": 82.5
},
{
"name": "minilm + 8-float head (2k labels)",
"acc": {
"sst2": 0.7972,
"ag_news": 0.8872,
"emotion": 0.6424,
"banking77": 0.8512
},
"lat": {
"sst2": 2.8222,
"ag_news": 2.8222,
"emotion": 2.8222,
"banking77": 2.8222
},
"ram": 451.7
},
{
"name": "mbert-embed + 8-float head (2k labels, proper score)",
"acc": {
"sst2": 0.876,
"ag_news": 0.8976,
"emotion": 0.6824,
"banking77": 0.8892
},
"lat": {
"sst2": 15.767900000000001,
"ag_news": 15.767900000000001,
"emotion": 15.767900000000001,
"banking77": 15.767900000000001
},
"ram": 489.5
}
],
"failures": [
[
"Template ensembling (3 NLI hypotheses)",
"0.0 pp",
"errors across templates are correlated"
],
[
"Evolutionary weight search, honest LODO",
"-1.5 pp",
"negative transfer; helps only when the base signal is weak"
],
[
"Cross-encoder 149M -> 395M",
"+2.8 / -3.2 / +0.2 pp",
"no net gain for 3x the compute"
],
[
"Listwise rerank with Qwen2.5-1.5B (top-5)",
"-1.0 pp",
"weaker than the cross-encoder alone"
],
[
"Column-centering NLI scores",
"0.0 pp macro",
"BANKING77 +2.6, Emotion -3.1"
],
[
"Asymmetric query/document prefixes",
"-1.8 pp",
"classification by similarity is not retrieval"
],
[
"Distil cross-encoder into bi-encoder head",
"-5.4 pp",
"collapses on Emotion (-18.8 pp)"
],
[
"Raw ModernBERT-base (MLM) as sentence encoder",
"35.0 %",
"worse than an 8M static embedding"
],
[
"ES objective = accuracy only",
"+1.0 pp when swapped",
"log score is a smoother signal; calibration was already fine"
]
],
"jev_cost_usd": 0.2841,
"jev_tokens": 6764108,
"extra": {
"laya_shortlist": {
"base": 0.5428,
"base_lat": 86.3,
"fixed": 0.608,
"fixed_lat": 39.8,
"ceiling": 0.9784,
"device": "mps",
"cpu": {
"base": 0.536,
"base_lat": 309.0,
"fixed": 0.62,
"fixed_lat": 169.9,
"n": 250,
"ceiling": 0.98
}
},
"chinese": {
"multilingual": {
"acc": 0.39166666666666666,
"conf": 0.7564826666666666,
"lat": 14.3
},
"english": {
"acc": 0.225,
"conf": 0.9681070000000002,
"lat": 42.5
},
"random": 0.06666666666666667
},
"objective": {
"acc": {
"per": {
"sst2": {
"acc": 0.858,
"ece": 0.013683630180358876,
"temp": 4.51,
"nll": 0.3143
},
"ag_news": {
"acc": 0.8948,
"ece": 0.03082646280527116,
"temp": 4.51,
"nll": 0.3369
},
"emotion": {
"acc": 0.6616,
"ece": 0.01806434515714644,
"temp": 4.87,
"nll": 0.9243
},
"banking77": {
"acc": 0.888,
"ece": 0.010833000487089142,
"temp": 3.59,
"nll": 0.384
}
},
"macro_acc": 0.8256,
"mean_ece": 0.018351859657466406,
"mean_nll": 0.48987499999999995,
"weights": {
"proto": 3.9147,
"desc": 0.8717,
"knn1": 2.5684,
"knnk": 2.8151,
"margin": 0.9435,
"prior": 0.3723,
"white": -0.9152,
"tight": 0.0329
}
},
"proper": {
"per": {
"sst2": {
"acc": 0.876,
"ece": 0.01733291234970094,
"temp": 5.25,
"nll": 0.2758
},
"ag_news": {
"acc": 0.8976,
"ece": 0.02284402102231979,
"temp": 4.18,
"nll": 0.3222
},
"emotion": {
"acc": 0.6824,
"ece": 0.021511850953102124,
"temp": 4.51,
"nll": 0.8761
},
"banking77": {
"acc": 0.8892,
"ece": 0.010253664314746886,
"temp": 3.33,
"nll": 0.375
}
},
"macro_acc": 0.8363,
"mean_ece": 0.017985612159967437,
"mean_nll": 0.462275,
"weights": {
"proto": 1.5202,
"desc": 0.8578,
"knn1": 1.6066,
"knnk": 2.7418,
"margin": 2.3915,
"prior": 0.4453,
"white": 0.4608,
"tight": -4.7245
}
}
},
"laya_cpu": {
"device": "cpu",
"n_per_dataset": 250,
"acc": {
"sst2": 0.856,
"ag_news": 0.948,
"emotion": 0.588,
"banking77": 0.536
},
"lat": {
"sst2": 59.5,
"ag_news": 78.0,
"emotion": 61.9,
"banking77": 309.0
},
"note": "Like-for-like with every other CPU row. ~3x the MPS latencies."
}
}
}