caden-encoder-mixed / evaluation-summary.json
Leonard02's picture
Release Caden v2: three-seed multi-domain continuation, calibration and audited accuracy/variance/speed
f1ab177 verified
Raw History Blame Contribute Delete
18.3 kB
{
"protocol": {
"version": "Caden-Multidomain-v2",
"seeds": [
42,
43,
44
],
"epochs": 1,
"lr": 1e-05,
"microbatch": 1,
"gradient_clip": 1,
"train_items": 12000,
"domain_counts": {
"banking": 3000,
"clinc": 3000,
"snips": 3000,
"agnews": 3000
},
"data_seed": 20261002,
"candidate_order_augmentation": true,
"checkpoint_initialization": "matching original mixed seed, not the exploratory pilot",
"calibration_items": 3996,
"new_target_calibration": 600,
"selection": "one pre-fixed epoch; release gate checks mean old/source dev retention (no worse than -1pp) and both target dev improvements. No test selection.",
"data_files": {
"agnews-dev.jsonl": {
"sha256": "935fed52715202b0ef173f7f4a4ae43a5b864c7c08967f72c28d08b716da63e2",
"bytes": 244615
},
"calibration.jsonl": {
"sha256": "0187dc4513c3af773e9ca5667e1e8144cf8d0593172353385ed4835c53528f2a",
"bytes": 2552849
},
"snips-dev.jsonl": {
"sha256": "ed984478535abb6cce639a2c53b455e60d68fd6f5b52197d464f346c35ec38fc",
"bytes": 354720
},
"source-dev.jsonl": {
"sha256": "7979b4e104ef8c5881989f373c10185fa5600728f1924524effd36ca69f82b43",
"bytes": 503568
},
"train.jsonl": {
"sha256": "d15909a3148ea50111848fb2bc9ae316edc2a2a1a5e62e985dff4137b99a375d",
"bytes": 7613311
}
},
"test_files": {
"banking": {
"path": "data/confirmation8-full/test.jsonl",
"sha256": "444a292af999764ff91c9be649b56a51633691f2fea967cca93196c1d89437b3"
},
"clinc": {
"path": "data/mixed-clinc-confirmation8/test.jsonl",
"sha256": "ad1275e418244db756424b0d83b8fa455a8c0841c26131d7536d97528bdbb099"
},
"snips": {
"path": "data/expansion-v1/snips/test.jsonl",
"sha256": "c922ce12a1395021570018c3276dd5872e96c4ddbbee73af63a3364d4cf13446"
},
"agnews": {
"path": "data/expansion-v1/agnews/test.jsonl",
"sha256": "ee73fe4bca74c900d367055b34ade57f44c92eb73218ff10923f18c53d84bdae"
}
},
"scope": "Continuation registered after exploratory prior benchmark reads; full test reruns are exploratory, not new untouched confirmations. Source replay and target supervision; no paid API. Old artifacts remain recoverable."
},
"training": [
{
"seed": 42,
"checkpoint": "checkpoints/caden-s42-multidomain-v2-e1",
"training_seconds": 291.6559239000053,
"dev": {
"source": {
"items": 800,
"failures": 0,
"accuracy_percent": 96.5,
"nll_raw_scores": 0.26655625034176156,
"latency_p50_ms": 3.341850002470892,
"latency_p95_ms": 4.701109996676678,
"latency_rate_per_second": 279.3293553582002,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"snips": {
"items": 700,
"failures": 0,
"accuracy_percent": 97.57142857142857,
"nll_raw_scores": 0.1774097527314909,
"latency_p50_ms": 3.2041499944170937,
"latency_p95_ms": 4.639800006407312,
"latency_rate_per_second": 296.31913641107684,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"agnews": {
"items": 400,
"failures": 0,
"accuracy_percent": 86.5,
"nll_raw_scores": 0.9300370775301481,
"latency_p50_ms": 3.208899994206149,
"latency_p95_ms": 4.2309099932026575,
"latency_rate_per_second": 302.19258859791603,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
}
}
},
{
"seed": 43,
"checkpoint": "checkpoints/caden-s43-multidomain-v2-e1",
"training_seconds": 289.73817129999225,
"dev": {
"source": {
"items": 800,
"failures": 0,
"accuracy_percent": 96.875,
"nll_raw_scores": 0.20278162164701882,
"latency_p50_ms": 3.449349998845719,
"latency_p95_ms": 5.495710007380694,
"latency_rate_per_second": 267.1525994275265,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"snips": {
"items": 700,
"failures": 0,
"accuracy_percent": 97.71428571428571,
"nll_raw_scores": 0.1518966859330835,
"latency_p50_ms": 3.273149995948188,
"latency_p95_ms": 4.524909992323955,
"latency_rate_per_second": 293.8327034137563,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"agnews": {
"items": 400,
"failures": 0,
"accuracy_percent": 86.5,
"nll_raw_scores": 0.7791759302059201,
"latency_p50_ms": 3.295449998404365,
"latency_p95_ms": 4.4259249945753245,
"latency_rate_per_second": 293.82108223346444,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
}
}
},
{
"seed": 44,
"checkpoint": "checkpoints/caden-s44-multidomain-v2-e1",
"training_seconds": 290.4029297000088,
"dev": {
"source": {
"items": 800,
"failures": 0,
"accuracy_percent": 97.0,
"nll_raw_scores": 0.18155284794806661,
"latency_p50_ms": 3.336500005389098,
"latency_p95_ms": 4.276090002531418,
"latency_rate_per_second": 293.0201465480304,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"snips": {
"items": 700,
"failures": 0,
"accuracy_percent": 97.14285714285714,
"nll_raw_scores": 0.2006310500109146,
"latency_p50_ms": 3.259199991589412,
"latency_p95_ms": 4.314464993512956,
"latency_rate_per_second": 296.28386802275867,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
},
"agnews": {
"items": 400,
"failures": 0,
"accuracy_percent": 88.25,
"nll_raw_scores": 0.8073049758331579,
"latency_p50_ms": 3.313449997222051,
"latency_p95_ms": 4.351695005607325,
"latency_rate_per_second": 292.45241007986886,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included"
}
}
}
],
"gates": {
"source_dev_change_pp": 0.9583333333333334,
"source_retention_passed": true,
"snips_dev_change_pp": 23.76190476190476,
"snips_improvement_passed": true,
"agnews_dev_change_pp": 38.83333333333333,
"agnews_improvement_passed": true
},
"tasks": {
"banking": {
"accuracy_mean_percent": 95.82115405434665,
"sample_sd_pp": 0.19844433019186825,
"sample_variance_pp_squared": 0.039380152185299236,
"seeds": [
42,
43,
44
],
"nll_raw_mean": 0.2941399471287365,
"calibrated_nll_mean": 0.16774011928050675,
"items": 3079,
"failures_total": 0
},
"clinc": {
"accuracy_mean_percent": 97.4823286474466,
"sample_sd_pp": 0.37848471312663967,
"sample_variance_pp_squared": 0.14325067807055472,
"seeds": [
42,
43,
44
],
"nll_raw_mean": 0.15676676241795667,
"calibrated_nll_mean": 0.09977191116593277,
"items": 4197,
"failures_total": 0
},
"snips": {
"accuracy_mean_percent": 97.04761904761905,
"sample_sd_pp": 0.5548181046538572,
"sample_variance_pp_squared": 0.3078231292516984,
"seeds": [
42,
43,
44
],
"nll_raw_mean": 0.2271088920342152,
"calibrated_nll_mean": 0.12925137847687118,
"items": 1400,
"failures_total": 0
},
"agnews": {
"accuracy_mean_percent": 87.42543859649122,
"sample_sd_pp": 0.13821009874200674,
"sample_variance_pp_squared": 0.01910203139427525,
"seeds": [
42,
43,
44
],
"nll_raw_mean": 0.8478021996165904,
"calibrated_nll_mean": 0.43472555140435853,
"items": 7600,
"failures_total": 0
}
},
"runs": [
{
"seed": 42,
"task": "banking",
"items": 3079,
"failures": 0,
"accuracy_percent": 95.64793764209159,
"nll_raw_scores": 0.33486586733779217,
"latency_p50_ms": 3.2648000051267445,
"latency_p95_ms": 4.279279998445417,
"latency_rate_per_second": 294.77747938070803,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.1955747003078493,
"prediction_sha256": "bdb607a47f0367bd4a40feec43e1fce1ec57e0071259ebd1188ff26cde164991"
},
{
"seed": 42,
"task": "clinc",
"items": 4197,
"failures": 0,
"accuracy_percent": 97.04550869668812,
"nll_raw_scores": 0.17161369503402035,
"latency_p50_ms": 3.192300006048754,
"latency_p95_ms": 4.240899998694657,
"latency_rate_per_second": 303.7195989965836,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.11655540477497105,
"prediction_sha256": "fb0f74b2fe2268fcce9e50081fe01c531fabff44056c9d82338f6f49e7efa5ca"
},
{
"seed": 42,
"task": "snips",
"items": 1400,
"failures": 0,
"accuracy_percent": 97.21428571428571,
"nll_raw_scores": 0.2255130143934313,
"latency_p50_ms": 3.1466500004171394,
"latency_p95_ms": 4.260794992296724,
"latency_rate_per_second": 306.42262701682176,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.12488741544444912,
"prediction_sha256": "952960d29aa68df9488950eb2f7380ffebbb194064a3d0a2bbf432be017d5c4b"
},
{
"seed": 42,
"task": "agnews",
"items": 7600,
"failures": 0,
"accuracy_percent": 87.5657894736842,
"nll_raw_scores": 0.8931826575703955,
"latency_p50_ms": 3.18400000105612,
"latency_p95_ms": 4.246409999905154,
"latency_rate_per_second": 304.13844625222157,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.43979012648828014,
"prediction_sha256": "8b9432e270f1776cee558e220eca21ae7b5d3eec6a23c59ae1e3aacaa14e4c1c"
},
{
"seed": 43,
"task": "banking",
"items": 3079,
"failures": 0,
"accuracy_percent": 96.03767456966547,
"nll_raw_scores": 0.2703388386872126,
"latency_p50_ms": 3.264099999796599,
"latency_p95_ms": 4.348850007227156,
"latency_rate_per_second": 297.59416814136756,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.15151505630540285,
"prediction_sha256": "10abe11ce72169434e4a97a5aa2f58e5ba92628f81d4e8c3f756b78821872138"
},
{
"seed": 43,
"task": "clinc",
"items": 4197,
"failures": 0,
"accuracy_percent": 97.71265189421015,
"nll_raw_scores": 0.13692723077173466,
"latency_p50_ms": 3.198200007318519,
"latency_p95_ms": 4.322980000870302,
"latency_rate_per_second": 301.69547683377635,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.0872236201855948,
"prediction_sha256": "d69542751eb015af8174856783d83dc9c110b11f520a7ba3d221bc803fc79e1f"
},
{
"seed": 43,
"task": "snips",
"items": 1400,
"failures": 0,
"accuracy_percent": 97.5,
"nll_raw_scores": 0.17975736809092224,
"latency_p50_ms": 3.157150000333786,
"latency_p95_ms": 4.312354991270695,
"latency_rate_per_second": 303.50836810969656,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.1085437458665211,
"prediction_sha256": "1648e54d36cdd77edc48d24bf3911437d54028ab67ae2459a4bba133fdef11bc"
},
{
"seed": 43,
"task": "agnews",
"items": 7600,
"failures": 0,
"accuracy_percent": 87.42105263157895,
"nll_raw_scores": 0.7655603057568001,
"latency_p50_ms": 3.279899996414315,
"latency_p95_ms": 4.559509998216527,
"latency_rate_per_second": 292.4148027308918,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.4163587879852833,
"prediction_sha256": "d6f75307e9a595c038cefa60c0e6a12c119f475a242ab72de417535d163ee661"
},
{
"seed": 44,
"task": "banking",
"items": 3079,
"failures": 0,
"accuracy_percent": 95.77784995128289,
"nll_raw_scores": 0.27721513536120473,
"latency_p50_ms": 3.3171999966725707,
"latency_p95_ms": 4.484859992226119,
"latency_rate_per_second": 292.1423167117449,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.15613060122826808,
"prediction_sha256": "86fbc02b69c62926824a4985d40e4498e0b82f8e1940f29242d4d401ae78d4bf"
},
{
"seed": 44,
"task": "clinc",
"items": 4197,
"failures": 0,
"accuracy_percent": 97.68882535144151,
"nll_raw_scores": 0.16175936144811504,
"latency_p50_ms": 3.222899991669692,
"latency_p95_ms": 4.257100005634129,
"latency_rate_per_second": 301.39727257581,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.09553670853723245,
"prediction_sha256": "d741e8fc1952dc2a8d3c7fcff0c1345698512fbf12c54296ab2351ac61aa56fc"
},
{
"seed": 44,
"task": "snips",
"items": 1400,
"failures": 0,
"accuracy_percent": 96.42857142857143,
"nll_raw_scores": 0.27605629361829204,
"latency_p50_ms": 3.1474499992327765,
"latency_p95_ms": 4.228460008016555,
"latency_rate_per_second": 307.54451058026063,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.15432297411964332,
"prediction_sha256": "4ea4dc0e8ffa67333f3cbec12feb7a714757ffb746ee9b66ffd2ef49bb65522e"
},
{
"seed": 44,
"task": "agnews",
"items": 7600,
"failures": 0,
"accuracy_percent": 87.28947368421052,
"nll_raw_scores": 0.8846636355225758,
"latency_p50_ms": 3.174500001478009,
"latency_p95_ms": 4.226320006273454,
"latency_rate_per_second": 305.18876148580233,
"temperature": 1.0,
"latency_scope": "FP32 eager GPU, two CPU threads, B1 tokenization+forward+scores readback+normalization+validation, no loading; raw scores readback included",
"calibrated_nll_raw_scores": 0.4480277397395121,
"prediction_sha256": "fe32eacc7b12767f1d3088bb5e32844a0ad83c74338efcbfa6d33971cd3036e3"
}
],
"timing_p50_ms": 3.1917500018607825,
"timing_latency_rate_per_second": 304.7388723282904,
"recovery": "All training succeeded. Original driver stopped while serializing a NumPy boundary boolean; scalar conversion fixed and calibration/evaluation resumed. Cached seed42 calibration reused; no retraining.",
"scope": "V2 exploratory multi-domain continuation; same fixed test tasks, recipes registered after earlier test reads. No causal architecture effect or pristine unseen-test claim."
}