laya-evaluation / banking_continuation_protocol.json
c41's picture
Publish evaluation preview and verified historical classification tables
ff51f29 verified
Raw History Blame Contribute Delete
1.76 kB
{
"models": [
{
"name": "historical",
"sha256": "a9bc527b2f95ec703ed38928ef05c75f6f6fabc56c4e2dfd4db82bc843d6b2e7"
},
{
"name": "control_13",
"sha256": "21382aca6ceb390637b38114dabaedbdb43118569d5fc0b0911c9123617aff3b"
},
{
"name": "control_42",
"sha256": "0fb9489f08c37e558f99f49b3b4690c40803f6563a141e7e01cc2d85f0255757"
},
{
"name": "control_87",
"sha256": "7154e60587bfb386c39f8e077ba29420ec6ced10e633d2020f081aae378a2c10"
},
{
"name": "sam_13",
"sha256": "0ff739f4413497119476af928b6cff482f1c0ef0cfe9bfb96527b769df172d29"
},
{
"name": "sam_42",
"sha256": "096594ba8e880e15c55a38897f80e19495a2e1d392ec092e2cd3bb3bc54fd8a4"
},
{
"name": "sam_87",
"sha256": "1a158c801472e7e8ba9db25d5bac33821a0a8a722fe47330f557cd9142df8808"
}
],
"source_sha256": "a288387babe8fbe45be9317aa7ef59066a31bdb97f7c82d48ed303b9edfc5131",
"precision": "eager FP32 SDPA",
"max_length": 128,
"batch_size": 32,
"selection": "all fixed final checkpoints; no selection",
"scope": "retrospective official test, previously observed; no subsequent tuning",
"bootstrap": "10000 shared row resamples; fixed checkpoints; excludes model-selection and seed population uncertainty",
"start_unix": 1791385111.935805,
"training": {
"epochs": 2,
"updates": 626,
"train_rows": 10003,
"former_dev_included_as_train": true,
"validation_selection": false,
"rho": 0.05,
"encoder_lr": 2e-06,
"head_lr": 2e-05,
"train_precision": "BF16",
"seeds": [
13,
42,
87
],
"initialization": "same historical full-train weights"
}
}