caden-encoder-mixed / release-protocol.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
2.33 kB
{
"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."
}