Manchego / eval /release_read.json
oraculumai's picture
Manchego v3
f82e029 verified
Raw History Blame Contribute Delete
10.5 kB
{
"what": "Manchego v3's release read: v3, Manchego v2.1 and the untrained Qwen3.5-4B scored in one runtime (bf16, one RTX 5090, padded batches of up to 8,192 tokens, the adapters applied to the base unmerged) on identical rows; aggregates only. JevBench numbers here are OUR runs on its 231 public decisions, not official JevBench scores.",
"names": {
"frozen (served)": "untrained Qwen3.5-4B under Manchego v2.1's served prompt policy (short prompt up to 26 options, state-first beyond)",
"frozen (SemIf)": "untrained Qwen3.5-4B under the SemIf prompt (state-first beyond 16 options)",
"Manchego v2.1 (served)": "Manchego v2.1 under its served prompt policy",
"v3m100b (SemIf)": "Manchego v3 (its adapter on the base, bf16) under the SemIf prompt (state-first beyond 16 options)"
},
"runtime_check_passed": true,
"runtime_check": "v2.1 and the untrained base reproduced their archived records on the three card suites (tolerance |accuracy| <= 0.01, |cross-entropy| <= 0.02)",
"intervals": "paired 95% percentile bootstraps; the resampling unit is named per suite ('interval')",
"card": {
"jevbench_public": {
"rows": 231,
"interval": "scenario groups resampled",
"systems": {
"frozen (served)": {
"n": 231,
"accuracy": 0.7748917748917749,
"cross_entropy": 0.49286616487181245,
"brier": 0.2866513305940945,
"ece": 0.06231002798067052,
"mean_confidence": 0.8204758784756144,
"refused": 0
},
"frozen (SemIf)": {
"n": 231,
"accuracy": 0.8008658008658008,
"cross_entropy": 0.4852084330739462,
"brier": 0.25419321047505944,
"ece": 0.0588310701620669,
"mean_confidence": 0.8366482323997403,
"refused": 0
},
"Manchego v2.1 (served)": {
"n": 231,
"accuracy": 0.8051948051948052,
"cross_entropy": 0.4232890730630488,
"brier": 0.24566343385005784,
"ece": 0.07188272887054453,
"mean_confidence": 0.8538814495170696,
"refused": 0
},
"v3m100b (SemIf)": {
"n": 231,
"accuracy": 0.8571428571428571,
"cross_entropy": 0.3260731101395454,
"brier": 0.18328252671634565,
"ece": 0.05534072233865077,
"mean_confidence": 0.8536973872249488,
"refused": 0
}
},
"v3_minus_v21": {
"accuracy": {
"mean": 0.05194805194805195,
"ci95": [
0.004201680672268907,
0.10222222222222223
]
},
"cross_entropy": {
"mean": -0.09721596292350353,
"ci95": [
-0.17117279532265608,
-0.03245969447592095
]
}
},
"v3_minus_frozen_served": {
"accuracy": {
"mean": 0.08225108225108226,
"ci95": [
0.021645021645021644,
0.14347826086956522
]
},
"cross_entropy": {
"mean": -0.16679305473226716,
"ci95": [
-0.2671700196376057,
-0.07280825498368075
]
}
}
},
"public8": {
"rows": 1600,
"interval": "items resampled within dataset",
"systems": {
"frozen (served)": {
"n": 1600,
"accuracy": 0.68625,
"cross_entropy": 0.8430894766320378,
"brier": 0.42806658962520244,
"ece": 0.12246943241697027,
"mean_confidence": 0.8087194324169702,
"refused": 0
},
"frozen (SemIf)": {
"n": 1600,
"accuracy": 0.725,
"cross_entropy": 0.8471491321273774,
"brier": 0.3999620136837791,
"ece": 0.12153813578564186,
"mean_confidence": 0.8439632063936646,
"refused": 0
},
"Manchego v2.1 (served)": {
"n": 1600,
"accuracy": 0.731875,
"cross_entropy": 0.6837753863639332,
"brier": 0.3671370658953875,
"ece": 0.05316138522715013,
"mean_confidence": 0.7847182733650019,
"refused": 0
},
"v3m100b (SemIf)": {
"n": 1600,
"accuracy": 0.753125,
"cross_entropy": 0.6085345175694903,
"brier": 0.3268992227572392,
"ece": 0.05543283372305037,
"mean_confidence": 0.8085578337230505,
"refused": 0
}
},
"v3_minus_v21": {
"accuracy": {
"mean": 0.02125,
"ci95": [
0.001875,
0.039375
]
},
"cross_entropy": {
"mean": -0.07524086879444299,
"ci95": [
-0.10700733270210035,
-0.04398069430837902
]
}
},
"v3_minus_frozen_served": {
"accuracy": {
"mean": 0.066875,
"ci95": [
0.04625,
0.08625
]
},
"cross_entropy": {
"mean": -0.23455495906254767,
"ci95": [
-0.27699129955904445,
-0.19314965712030813
]
}
}
},
"transfer": {
"rows": 1561,
"interval": "the 27 source-clean tasks (v2.1 card), tasks resampled",
"systems": {
"frozen (served)": {
"n": 1561,
"accuracy": 0.6777706598334401,
"cross_entropy": 0.9189782184185465,
"brier": 0.43554421632973334,
"ece": 0.10705508985338476,
"mean_confidence": 0.7840551579972183,
"refused": 0
},
"frozen (SemIf)": {
"n": 1561,
"accuracy": 0.6777706598334401,
"cross_entropy": 1.055959778577908,
"brier": 0.46561897696066334,
"ece": 0.11924407258759004,
"mean_confidence": 0.7967923988846295,
"refused": 0
},
"Manchego v2.1 (served)": {
"n": 1561,
"accuracy": 0.7014734144778988,
"cross_entropy": 0.9285998666239199,
"brier": 0.42428692850163635,
"ece": 0.08289752004742494,
"mean_confidence": 0.7837359571214004,
"refused": 0
},
"v3m100b (SemIf)": {
"n": 1561,
"accuracy": 0.7085201793721974,
"cross_entropy": 0.9408154991498171,
"brier": 0.4318082093390665,
"ece": 0.12920627298362403,
"mean_confidence": 0.8292848345361473,
"refused": 0
}
},
"v3_minus_v21": {
"accuracy": {
"mean": 0.007046764894298526,
"ci95": [
-0.01585288522511097,
0.03132992327365729
]
},
"cross_entropy": {
"mean": 0.012215632525897134,
"ci95": [
-0.20211863222827312,
0.15664732848488094
]
}
},
"v3_minus_frozen_served": {
"accuracy": {
"mean": 0.03074951953875721,
"ci95": [
0.0019157088122605363,
0.06482670089858793
]
},
"cross_entropy": {
"mean": 0.02183728073127055,
"ci95": [
-0.08835037289694243,
0.15047515066869643
]
}
}
}
},
"sealed": {
"confirm_b": {
"what": "24 Natural Instructions tasks from 13 source clusters no Manchego model trained on, read once; the contrast is the mean over tasks of the per-task accuracy difference, source clusters resampled; 'systems' are pooled over rows",
"claim": {
"v3_minus_v21": {
"tasks": 24,
"clusters": 13,
"rows": 5397,
"accuracy": {
"mean": 0.04227827381234498,
"ci95": [
-0.017932316870179994,
0.1068637823646576
]
},
"cross_entropy": {
"mean": 0.0024332706544883936,
"ci95": [
-0.10795744285557755,
0.1274440740433659
]
}
},
"v3_minus_frozen_served": {
"tasks": 24,
"clusters": 13,
"rows": 5397,
"accuracy": {
"mean": 0.05955055602143098,
"ci95": [
-0.019663513011422513,
0.14306252573617587
]
},
"cross_entropy": {
"mean": -0.055827519937997915,
"ci95": [
-0.15976968646042974,
0.07125067495963548
]
}
},
"systems": {
"frozen (served)": {
"n": 5397,
"accuracy": 0.6409116175653141,
"cross_entropy": 0.9178368709979563,
"brier": 0.48264586072708016,
"ece": 0.15134806828254388,
"mean_confidence": 0.7922596858478579,
"refused": 0
},
"frozen (SemIf)": {
"n": 5397,
"accuracy": 0.6740781915879193,
"cross_entropy": 0.9107246561067318,
"brier": 0.46026264250869503,
"ece": 0.1475488337401401,
"mean_confidence": 0.8213640799547319,
"refused": 0
},
"Manchego v2.1 (served)": {
"n": 5397,
"accuracy": 0.6588845654993515,
"cross_entropy": 0.8486214459422593,
"brier": 0.46015495267198037,
"ece": 0.09823534911045727,
"mean_confidence": 0.7571199146098088,
"refused": 0
},
"v3m100b (SemIf)": {
"n": 5397,
"accuracy": 0.6950157494904576,
"cross_entropy": 0.8607399041168844,
"brier": 0.44547195424681324,
"ece": 0.14094955334255857,
"mean_confidence": 0.8359653028330162,
"refused": 0
}
}
}
},
"rep_sealed": {
"what": "fresh out-of-distribution draws of four of the project's own computed families (policy decisions, lookup chains, temporal, answer adequacy), 400 instances, read once; instances resampled",
"rows": 1193,
"instances": 400,
"systems": {
"frozen (served)": {
"n": 1193,
"accuracy": 0.4450963956412406,
"cross_entropy": 1.1611160039257353,
"brier": 0.6460967922428869,
"ece": 0.1405380034173992,
"mean_confidence": 0.583342473485591,
"refused": 0
},
"frozen (SemIf)": {
"n": 1193,
"accuracy": 0.5079631181894384,
"cross_entropy": 1.0912352055652879,
"brier": 0.5978876795541787,
"ece": 0.07519011377082045,
"mean_confidence": 0.5831532319602589,
"refused": 0
},
"Manchego v2.1 (served)": {
"n": 1193,
"accuracy": 0.7946353730092205,
"cross_entropy": 0.4659314334385817,
"brier": 0.2733922531947281,
"ece": 0.051271641242572615,
"mean_confidence": 0.8437516230506135,
"refused": 0
},
"v3m100b (SemIf)": {
"n": 1193,
"accuracy": 0.7619446772841576,
"cross_entropy": 0.5777628650382641,
"brier": 0.3190441194412629,
"ece": 0.0495448301892173,
"mean_confidence": 0.8004757787551626,
"refused": 0
}
},
"v3_minus_v21": {
"accuracy": {
"mean": -0.03269069572506287,
"ci95": [
-0.060085836909871244,
-0.005106382978723404
]
},
"cross_entropy": {
"mean": 0.11183143159968244,
"ci95": [
0.04482470675698149,
0.18231048096005373
]
}
},
"v3_minus_frozen_served": {
"accuracy": {
"mean": 0.316848281642917,
"ci95": [
0.2802013422818792,
0.35313807531380753
]
},
"cross_entropy": {
"mean": -0.5833531388874713,
"ci95": [
-0.6449727284665607,
-0.519783716913897
]
}
}
}
},
"flags": {
"runtime_check_passed": true,
"jevbench_v3_below_v21_point": false,
"resolved_below_v21": [
"rep_sealed"
]
}
}