Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibration-budget/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores/immicro/rankcal_20261001.json

@@ -0,0 +1,641 @@
+{
+ "lora_r4_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.16270667791846324,
+    0.03727093052841324,
+    5
+   ],
+   "precision": [
+    0.9895584415584416,
+    0.0014417599276791112,
+    5
+   ],
+   "coverage": [
+    0.9072323232323232,
+    0.026518561294827336,
+    5
+   ],
+   "pb_r_median": [
+    0.9021729063620997,
+    0.0032708643327212963,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.16221250387191163,
+   "precision": 0.9368847352024922,
+   "coverage": 0.7890965732087227,
+   "pb_r_median": 0.9023670557055304,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    3750,
+    3250,
+    3250,
+    16750,
+    4500
+   ],
+   "inner_val": 0.10287906910691942,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_r4_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.1721542914277937,
+    0.020397497521952815,
+    5
+   ],
+   "precision": [
+    0.9859336219336221,
+    0.006439808435291671,
+    5
+   ],
+   "coverage": [
+    0.8743520923520922,
+    0.014800248956596629,
+    5
+   ],
+   "pb_r_median": [
+    0.8995046652352034,
+    0.0009659021499981422,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.17095159865002751,
+   "precision": 0.9247352024922119,
+   "coverage": 0.7149532710280374,
+   "pb_r_median": 0.8993883672634128,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    13000,
+    5500,
+    20000,
+    19750,
+    6250
+   ],
+   "inner_val": 0.10238224683063371,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_r4__raw": {
+  "fold": {
+   "comp_l1": [
+    0.18499427015987308,
+    0.02175014552772835,
+    5
+   ],
+   "precision": [
+    0.9887878787878787,
+    0.002490337893772557,
+    5
+   ],
+   "coverage": [
+    0.8770072150072149,
+    0.02905764520958238,
+    5
+   ],
+   "pb_r_median": [
+    0.8989815470414305,
+    0.002313639924694424,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.18577808236758375,
+   "precision": 0.9333333333333333,
+   "coverage": 0.7123364485981308,
+   "pb_r_median": 0.8991863255707665,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    14750,
+    15250,
+    20000,
+    18250,
+    18250
+   ],
+   "inner_val": 0.10214247618402754,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.17044981248033717,
+    0.027120470271117504,
+    5
+   ],
+   "precision": [
+    0.9901471861471862,
+    0.0017588761103520725,
+    5
+   ],
+   "coverage": [
+    0.9270966810966812,
+    0.032279442998871886,
+    5
+   ],
+   "pb_r_median": [
+    0.9046183714227549,
+    0.004464428846410777,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.17145118443734877,
+   "precision": 0.9462928348909657,
+   "coverage": 0.839190031152648,
+   "pb_r_median": 0.9045080329003975,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    3250,
+    2750,
+    3000,
+    5000,
+    3500
+   ],
+   "inner_val": 0.10280675568750927,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "lora_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.17076937636118156,
+    0.020828812823069094,
+    5
+   ],
+   "precision": [
+    0.986992784992785,
+    0.003724280216945599,
+    5
+   ],
+   "coverage": [
+    0.8983463203463204,
+    0.015144537663637428,
+    5
+   ],
+   "pb_r_median": [
+    0.8990964586739695,
+    0.0035244935725606994,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.17232428387355106,
+   "precision": 0.9204361370716511,
+   "coverage": 0.755202492211838,
+   "pb_r_median": 0.8995000568389205,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    19750,
+    19500,
+    20000,
+    19500,
+    5250
+   ],
+   "inner_val": 0.10208792324577058,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "real_lora__raw": {
+  "fold": {
+   "comp_l1": [
+    0.1617283055645703,
+    0.016796626065918724,
+    5
+   ],
+   "precision": [
+    0.9886262626262626,
+    0.0027380244179385886,
+    5
+   ],
+   "coverage": [
+    0.9185165945165945,
+    0.008612599776653159,
+    5
+   ],
+   "pb_r_median": [
+    0.8995792807263573,
+    0.0011744399503695433,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.16172890350786212,
+   "precision": 0.9241121495327103,
+   "coverage": 0.7842990654205607,
+   "pb_r_median": 0.8994149755898816,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    20000,
+    19500,
+    20000,
+    20000,
+    20000
+   ],
+   "inner_val": 0.10171123402459281,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "lora_r64_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.17626059507375783,
+    0.02731799096341323,
+    5
+   ],
+   "precision": [
+    0.990836940836941,
+    0.0017005390894841556,
+    5
+   ],
+   "coverage": [
+    0.9267070707070706,
+    0.036294154741976,
+    5
+   ],
+   "pb_r_median": [
+    0.903582858682092,
+    0.003971821399560073,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.17651083150222252,
+   "precision": 0.9455451713395638,
+   "coverage": 0.8472897196261682,
+   "pb_r_median": 0.9035992444051067,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    2750,
+    2500,
+    2500,
+    3250,
+    2750
+   ],
+   "inner_val": 0.10277071190731865,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "lora_r64_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.17098847953604462,
+    0.021576704468132847,
+    5
+   ],
+   "precision": [
+    0.9838759018759019,
+    0.004088835682447718,
+    5
+   ],
+   "coverage": [
+    0.9182337662337663,
+    0.021640888494273102,
+    5
+   ],
+   "pb_r_median": [
+    0.8984489965954143,
+    0.003537793749598114,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.1716537920789307,
+   "precision": 0.9052959501557633,
+   "coverage": 0.7937071651090343,
+   "pb_r_median": 0.8986129241189431,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    20000,
+    19750,
+    20000,
+    20000,
+    3750
+   ],
+   "inner_val": 0.10185143617647034,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "lora_r64__raw": {
+  "fold": {
+   "comp_l1": [
+    0.16099156472072332,
+    0.01654368259136177,
+    5
+   ],
+   "precision": [
+    0.9853852813852815,
+    0.0040537734898083155,
+    5
+   ],
+   "coverage": [
+    0.94001443001443,
+    0.007887742085254048,
+    5
+   ],
+   "pb_r_median": [
+    0.899712035989664,
+    0.0014595430187922984,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.1609507558094419,
+   "precision": 0.9044236760124611,
+   "coverage": 0.8158878504672897,
+   "pb_r_median": 0.8995201341884855,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    20000,
+    19750,
+    20000,
+    20000,
+    20000
+   ],
+   "inner_val": 0.1012812223817621,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n5": {
+  "fold": {
+   "comp_l1": [
+    0.19112194334505275,
+    0.029738749755902644,
+    5
+   ],
+   "precision": [
+    0.977774891774892,
+    0.004651172182618444,
+    5
+   ],
+   "coverage": [
+    0.7255411255411257,
+    0.00993146201782128,
+    5
+   ],
+   "pb_r_median": [
+    0.8804044974427396,
+    0.002020094647805375,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.19164257165311652,
+   "precision": 0.9035514018691588,
+   "coverage": 0.5270404984423676,
+   "pb_r_median": 0.8806309938606883,
+   "deg_top50_sum": null
+  }
+ },
+ "lora_n10__raw": {
+  "fold": {
+   "comp_l1": [
+    0.18767172159138049,
+    0.012004033609613402,
+    5
+   ],
+   "precision": [
+    0.9892958152958153,
+    0.005256175095229,
+    5
+   ],
+   "coverage": [
+    0.898060606060606,
+    0.039550382847834395,
+    5
+   ],
+   "pb_r_median": [
+    0.9003594475037394,
+    0.0028568719301180355,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.18788133961403,
+   "precision": 0.9381308411214954,
+   "coverage": 0.7856697819314642,
+   "pb_r_median": 0.9009528958270383,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    5750,
+    3250,
+    4000,
+    17000,
+    19500
+   ],
+   "inner_val": 0.10238993998084749,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n10": {
+  "fold": {
+   "comp_l1": [
+    0.20079865322750887,
+    0.025468178276030453,
+    5
+   ],
+   "precision": [
+    0.9777575757575757,
+    0.0034299692031249866,
+    5
+   ],
+   "coverage": [
+    0.7276277056277055,
+    0.00696970892000926,
+    5
+   ],
+   "pb_r_median": [
+    0.8799493297830476,
+    0.0022637239366329748,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.1996025683151641,
+   "precision": 0.9066043613707165,
+   "coverage": 0.5287227414330218,
+   "pb_r_median": 0.8799329202328436,
+   "deg_top50_sum": null
+  }
+ },
+ "zs__cal_n20": {
+  "fold": {
+   "comp_l1": [
+    0.19182606516653886,
+    0.027852292949926565,
+    5
+   ],
+   "precision": [
+    0.9777777777777776,
+    0.0032669406074554595,
+    5
+   ],
+   "coverage": [
+    0.721027417027417,
+    0.007531606381806122,
+    5
+   ],
+   "pb_r_median": [
+    0.8801365953826066,
+    0.0024909162630608744,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.19112138214206625,
+   "precision": 0.910404984423676,
+   "coverage": 0.5287227414330218,
+   "pb_r_median": 0.8804933305805398,
+   "deg_top50_sum": null
+  }
+ },
+ "lora_n40__raw": {
+  "fold": {
+   "comp_l1": [
+    0.17078367214496734,
+    0.009478376690859469,
+    5
+   ],
+   "precision": [
+    0.9875699855699855,
+    0.0022419896284439,
+    5
+   ],
+   "coverage": [
+    0.9138383838383838,
+    0.01879281005809579,
+    5
+   ],
+   "pb_r_median": [
+    0.8989409274301723,
+    0.002160110672595629,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.17138488720357878,
+   "precision": 0.9229283489096574,
+   "coverage": 0.7784423676012461,
+   "pb_r_median": 0.8990074307741415,
+   "deg_top50_sum": null
+  },
+  "train": {
+   "selected_steps": [
+    20000,
+    19500,
+    20000,
+    20000,
+    20000
+   ],
+   "inner_val": 0.10182350309831756,
+   "pretrained_inner_val": 0.11045852335436004,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n40": {
+  "fold": {
+   "comp_l1": [
+    0.19560956823822193,
+    0.025815798489759543,
+    5
+   ],
+   "precision": [
+    0.9764242424242424,
+    0.005711610433847429,
+    5
+   ],
+   "coverage": [
+    0.7336190476190476,
+    0.009109925368529502,
+    5
+   ],
+   "pb_r_median": [
+    0.880179996909769,
+    0.002482781542820341,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.19600612924651933,
+   "precision": 0.9077258566978194,
+   "coverage": 0.5223676012461059,
+   "pb_r_median": 0.8803800757617266,
+   "deg_top50_sum": null
+  }
+ },
+ "zs__cal": {
+  "fold": {
+   "comp_l1": [
+    0.19904720741429882,
+    0.02892928783568735,
+    5
+   ],
+   "precision": [
+    0.9760606060606062,
+    0.006076149928907217,
+    5
+   ],
+   "coverage": [
+    0.7254516594516595,
+    0.006451526962691748,
+    5
+   ],
+   "pb_r_median": [
+    0.879899828601547,
+    0.002578710363793526,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.19850275515696958,
+   "precision": 0.9050467289719626,
+   "coverage": 0.527227414330218,
+   "pb_r_median": 0.8801553120873381,
+   "deg_top50_sum": null
+  }
+ }
+}

paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibration-budget/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores/immicro/rankcal_20261001.md

@@ -0,0 +1,34 @@
+# immicro: LoRA rank sweep and equal-budget calibration (2026-10-01)
+
+Existing arms: scores/immicro/fold<k>, scores/immicro/pooled; new: scores/immicro/lora_rank_20261001, scores/immicro/cal_budget_20261001.
+
+## LoRA rank (alpha = r; comparator = r 16 at the same n)
+
+| n | r | comp L1 fold mean ± SD (folds) | coverage (fold mean) | pooled comp L1 | pooled precision / coverage | selected step; inner-val loss | pooled per-donor L1 vs comparator (Δ, better in, Wilcoxon p) |
+|---|---|---|---|---|---|---|---|
+| 5 | 4 | 0.163 ± 0.037 (5) | 0.907 | 0.162 | 0.937 / 0.789 | 3250–16750; 0.10288 | -0.009 (64% better, p 0.00088) |
+| 5 | 16 | 0.170 ± 0.027 (5) | 0.927 | 0.171 | 0.946 / 0.839 | 2750–5000; 0.10281 | (comparator) |
+| 5 | 64 | 0.176 ± 0.027 (5) | 0.927 | 0.177 | 0.946 / 0.847 | 2500–3250; 0.10277 | +0.005 (40% better, p 0.0076) |
+| 20 | 4 | 0.172 ± 0.020 (5) | 0.874 | 0.171 | 0.925 / 0.715 | 5500–20000; 0.10238 | -0.001 (54% better, p 0.32) |
+| 20 | 16 | 0.171 ± 0.021 (5) | 0.898 | 0.172 | 0.920 / 0.755 | 5250–20000; 0.10209 | (comparator) |
+| 20 | 64 | 0.171 ± 0.022 (5) | 0.918 | 0.172 | 0.905 / 0.794 | 3750–20000; 0.10185 | -0.001 (44% better, p 0.66) |
+| all | 4 | 0.185 ± 0.022 (5) | 0.877 | 0.186 | 0.933 / 0.712 | 14750–20000; 0.10214 | +0.024 (27% better, p 9.8e-09) |
+| all | 16 | 0.162 ± 0.017 (5) | 0.919 | 0.162 | 0.924 / 0.784 | 19500–20000; 0.10171 | (comparator) |
+| all | 64 | 0.161 ± 0.017 (5) | 0.940 | 0.161 | 0.904 / 0.816 | 19750–20000; 0.10128 | -0.001 (51% better, p 0.89) |
+
+## Zero-shot paired calibration at the LoRA pair budget (comparator = zs__cal_n<n>)
+
+| n | arm | comp L1 fold mean ± SD (folds) | coverage (fold mean) | pooled comp L1 | pooled precision / coverage | selected step; inner-val loss | pooled per-donor L1 vs comparator (Δ, better in, Wilcoxon p) |
+|---|---|---|---|---|---|---|---|
+| 5 | zs__cal_n5 | 0.191 ± 0.030 (5) | 0.726 | 0.192 | 0.904 / 0.527 |  | (comparator) |
+| 5 | lora_n5__raw | 0.170 ± 0.027 (5) | 0.927 | 0.171 | 0.946 / 0.839 | 2750–5000; 0.10281 | -0.020 (62% better, p 0.00078) |
+| 10 | zs__cal_n10 | 0.201 ± 0.025 (5) | 0.728 | 0.200 | 0.907 / 0.529 |  | (comparator) |
+| 10 | lora_n10__raw | 0.188 ± 0.012 (5) | 0.898 | 0.188 | 0.938 / 0.786 | 3250–19500; 0.10239 | -0.012 (63% better, p 0.03) |
+| 20 | zs__cal_n20 | 0.192 ± 0.028 (5) | 0.721 | 0.191 | 0.910 / 0.529 |  | (comparator) |
+| 20 | lora_n20__raw | 0.171 ± 0.021 (5) | 0.898 | 0.172 | 0.920 / 0.755 | 5250–20000; 0.10209 | -0.019 (65% better, p 0.00083) |
+| 40 | zs__cal_n40 | 0.196 ± 0.026 (5) | 0.734 | 0.196 | 0.908 / 0.522 |  | (comparator) |
+| 40 | lora_n40__raw | 0.171 ± 0.009 (5) | 0.914 | 0.171 | 0.923 / 0.778 | 19500–20000; 0.10182 | -0.025 (69% better, p 0.00022) |
+| all | zs__cal | 0.199 ± 0.029 (5) | 0.725 | 0.199 | 0.905 / 0.527 |  | (comparator) |
+| all | real_lora__raw | 0.162 ± 0.017 (5) | 0.919 | 0.162 | 0.924 / 0.784 | 19500–20000; 0.10171 | -0.037 (69% better, p 1.9e-07) |
+
+Fold scores use each fold's test donors as reference; pooled scores use all donors (cross-fitted). n = training pairs; every run also saw the fold's inner-validation pairs, which zs__cal_n<n> uses for alpha.

paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibration-budget/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/rankcal_20261001.json

@@ -0,0 +1,673 @@
+{
+ "lora_r4_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.42713668090237,
+    0.05986485945097205,
+    5
+   ],
+   "precision": [
+    0.9872500000000001,
+    0.005056996839222948,
+    5
+   ],
+   "coverage": [
+    0.7800637254901961,
+    0.04330958156510056,
+    5
+   ],
+   "pb_r_median": [
+    0.8773806445286889,
+    0.01032289387488151,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.42606212341964195,
+   "precision": 0.9065853658536586,
+   "coverage": 0.7419512195121951,
+   "pb_r_median": 0.8766667848371756,
+   "deg_top50_sum": 2,
+   "logfc_r": 0.272342903622687,
+   "partial_r": 0.2318532187868404
+  },
+  "train": {
+   "selected_steps": [
+    2750,
+    2250,
+    3000,
+    2000,
+    3500
+   ],
+   "inner_val": 0.11271696762580956,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_r4_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.37877045909610824,
+    0.07542286315976757,
+    5
+   ],
+   "precision": [
+    0.9900637254901961,
+    0.004195853656706441,
+    5
+   ],
+   "coverage": [
+    0.832578431372549,
+    0.04581696782373666,
+    5
+   ],
+   "pb_r_median": [
+    0.8866751804539927,
+    0.004917414627138197,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.3777491255636951,
+   "precision": 0.9290243902439025,
+   "coverage": 0.7691056910569106,
+   "pb_r_median": 0.8878055061349732,
+   "deg_top50_sum": 14,
+   "logfc_r": 0.3122090069558179,
+   "partial_r": 0.23902344509543627
+  },
+  "train": {
+   "selected_steps": [
+    3250,
+    3500,
+    3750,
+    2750,
+    4000
+   ],
+   "inner_val": 0.1119349189368742,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_r4__raw": {
+  "fold": {
+   "comp_l1": [
+    0.3648133325241136,
+    0.06817373419708458,
+    5
+   ],
+   "precision": [
+    0.9883823529411766,
+    0.0026073293943324285,
+    5
+   ],
+   "coverage": [
+    0.8554558823529412,
+    0.0323061000410981,
+    5
+   ],
+   "pb_r_median": [
+    0.8901418019474198,
+    0.003655536309893239,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.3634628129626395,
+   "precision": 0.9232520325203252,
+   "coverage": 0.7887804878048781,
+   "pb_r_median": 0.8906670659863273,
+   "deg_top50_sum": 18,
+   "logfc_r": 0.4762839576698495,
+   "partial_r": 0.4157524998071602
+  },
+  "train": {
+   "selected_steps": [
+    3750,
+    4750,
+    6750,
+    10500,
+    7000
+   ],
+   "inner_val": 0.11158845073410442,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 324056,
+   "n_runs": 5
+  }
+ },
+ "lora_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.43017513856387357,
+    0.07442087038402356,
+    5
+   ],
+   "precision": [
+    0.9841862745098038,
+    0.007352581690558771,
+    5
+   ],
+   "coverage": [
+    0.7849852941176471,
+    0.03646646020808232,
+    5
+   ],
+   "pb_r_median": [
+    0.8773166085972444,
+    0.010114172298871075,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.4280613786711354,
+   "precision": 0.9114634146341464,
+   "coverage": 0.7666666666666667,
+   "pb_r_median": 0.8773560157667726,
+   "deg_top50_sum": 5,
+   "logfc_r": 0.29404452074006043,
+   "partial_r": 0.25032101587181504
+  },
+  "train": {
+   "selected_steps": [
+    2750,
+    2000,
+    2750,
+    2000,
+    3000
+   ],
+   "inner_val": 0.11266283636380518,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "lora_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.3817361394711213,
+    0.07942157721806865,
+    5
+   ],
+   "precision": [
+    0.9882450980392157,
+    0.004125284700018801,
+    5
+   ],
+   "coverage": [
+    0.8407598039215687,
+    0.04644478502001657,
+    5
+   ],
+   "pb_r_median": [
+    0.8849127180421845,
+    0.005911303949576034,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.3802570478327888,
+   "precision": 0.9266666666666666,
+   "coverage": 0.8066666666666666,
+   "pb_r_median": 0.8861159890865429,
+   "deg_top50_sum": 18,
+   "logfc_r": 0.3671933060907342,
+   "partial_r": 0.294607811871337
+  },
+  "train": {
+   "selected_steps": [
+    3000,
+    3000,
+    3250,
+    2500,
+    4000
+   ],
+   "inner_val": 0.1118740816866713,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "real_lora__raw": {
+  "fold": {
+   "comp_l1": [
+    0.36313994927777793,
+    0.07153269983249065,
+    5
+   ],
+   "precision": [
+    0.9880539215686275,
+    0.002650098109940577,
+    5
+   ],
+   "coverage": [
+    0.8648137254901961,
+    0.03270180392736473,
+    5
+   ],
+   "pb_r_median": [
+    0.8902277822907216,
+    0.004092736443911285,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.3618246783833512,
+   "precision": 0.9256910569105691,
+   "coverage": 0.829349593495935,
+   "pb_r_median": 0.890814724779387,
+   "deg_top50_sum": 23,
+   "logfc_r": 0.5267612711229187,
+   "partial_r": 0.46874787884206276
+  },
+  "train": {
+   "selected_steps": [
+    6500,
+    3750,
+    13000,
+    6250,
+    6750
+   ],
+   "inner_val": 0.1114082874198045,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "lora_r64_n5__raw": {
+  "fold": {
+   "comp_l1": [
+    0.4316734194717874,
+    0.09858166706568042,
+    5
+   ],
+   "precision": [
+    0.9842892156862746,
+    0.008801408818671416,
+    5
+   ],
+   "coverage": [
+    0.783357843137255,
+    0.04752808059272867,
+    5
+   ],
+   "pb_r_median": [
+    0.8779814537740659,
+    0.009624276211136878,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.42930844877241436,
+   "precision": 0.9064227642276422,
+   "coverage": 0.7660975609756098,
+   "pb_r_median": 0.8775560117022416,
+   "deg_top50_sum": 1,
+   "logfc_r": 0.30598162082894986,
+   "partial_r": 0.26099373757581557
+  },
+  "train": {
+   "selected_steps": [
+    2500,
+    2000,
+    2500,
+    2000,
+    2750
+   ],
+   "inner_val": 0.11259006746113301,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "lora_r64_n20__raw": {
+  "fold": {
+   "comp_l1": [
+    0.37813966580850983,
+    0.07542264123004411,
+    5
+   ],
+   "precision": [
+    0.9851666666666666,
+    0.009367225726973247,
+    5
+   ],
+   "coverage": [
+    0.8396078431372549,
+    0.04675076288937087,
+    5
+   ],
+   "pb_r_median": [
+    0.8869202712104451,
+    0.004031520791198365,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.37637955977000254,
+   "precision": 0.9105691056910569,
+   "coverage": 0.8079674796747968,
+   "pb_r_median": 0.8873403937326589,
+   "deg_top50_sum": 18,
+   "logfc_r": 0.43036005920508624,
+   "partial_r": 0.357485759477321
+  },
+  "train": {
+   "selected_steps": [
+    2750,
+    2500,
+    4000,
+    3750,
+    3500
+   ],
+   "inner_val": 0.1118157263046929,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "lora_r64__raw": {
+  "fold": {
+   "comp_l1": [
+    0.36856688183309644,
+    0.055774541941903726,
+    5
+   ],
+   "precision": [
+    0.9827990196078431,
+    0.0040022491033703195,
+    5
+   ],
+   "coverage": [
+    0.8693872549019609,
+    0.03789663526537358,
+    5
+   ],
+   "pb_r_median": [
+    0.8898874767056162,
+    0.003298904388460317,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.36687701383619603,
+   "precision": 0.9071544715447154,
+   "coverage": 0.8244715447154471,
+   "pb_r_median": 0.8909030338123556,
+   "deg_top50_sum": 26,
+   "logfc_r": 0.5529366684804276,
+   "partial_r": 0.49205208323159033
+  },
+  "train": {
+   "selected_steps": [
+    3750,
+    19500,
+    5000,
+    5750,
+    4250
+   ],
+   "inner_val": 0.11128811313371574,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 5184896,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n5": {
+  "fold": {
+   "comp_l1": [
+    0.413158422179274,
+    0.0794966637725908,
+    5
+   ],
+   "precision": [
+    0.987078431372549,
+    0.004596793959476799,
+    5
+   ],
+   "coverage": [
+    0.6481813725490196,
+    0.054155093697590875,
+    5
+   ],
+   "pb_r_median": [
+    0.8595459504444115,
+    0.004334499364402429,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.4119864488366454,
+   "precision": 0.898130081300813,
+   "coverage": 0.4353658536585366,
+   "pb_r_median": 0.8606745709395913,
+   "deg_top50_sum": 12,
+   "logfc_r": 0.07938905362315378,
+   "partial_r": -0.015295574290050627
+  }
+ },
+ "lora_n10__raw": {
+  "fold": {
+   "comp_l1": [
+    0.3921065176291368,
+    0.060375047253635836,
+    5
+   ],
+   "precision": [
+    0.9878921568627451,
+    0.0016885992012882602,
+    5
+   ],
+   "coverage": [
+    0.8238872549019607,
+    0.030984408673303647,
+    5
+   ],
+   "pb_r_median": [
+    0.8825504341742798,
+    0.0061314417115708635,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.38966522903618334,
+   "precision": 0.916260162601626,
+   "coverage": 0.7858536585365854,
+   "pb_r_median": 0.8822561969335914,
+   "deg_top50_sum": 8,
+   "logfc_r": 0.4222959020688887,
+   "partial_r": 0.36904669114334154
+  },
+  "train": {
+   "selected_steps": [
+    3250,
+    2000,
+    3000,
+    2250,
+    3000
+   ],
+   "inner_val": 0.11226841059646438,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n10": {
+  "fold": {
+   "comp_l1": [
+    0.4153228942393866,
+    0.07840882595961508,
+    5
+   ],
+   "precision": [
+    0.9888970588235294,
+    0.002116020957918022,
+    5
+   ],
+   "coverage": [
+    0.6455392156862745,
+    0.047810439018107644,
+    5
+   ],
+   "pb_r_median": [
+    0.8600845194334467,
+    0.004427094532326494,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.4139614715438767,
+   "precision": 0.8960975609756098,
+   "coverage": 0.44146341463414634,
+   "pb_r_median": 0.8609571113749166,
+   "deg_top50_sum": 10,
+   "logfc_r": 0.09542918105844352,
+   "partial_r": 0.004585854385854624
+  }
+ },
+ "zs__cal_n20": {
+  "fold": {
+   "comp_l1": [
+    0.4089407433506718,
+    0.07708664363535796,
+    5
+   ],
+   "precision": [
+    0.9882696078431372,
+    0.00394423596007229,
+    5
+   ],
+   "coverage": [
+    0.6370735294117648,
+    0.046621157163578106,
+    5
+   ],
+   "pb_r_median": [
+    0.8604460328542111,
+    0.0042661224674075526,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.40747903968889737,
+   "precision": 0.8992682926829269,
+   "coverage": 0.42520325203252035,
+   "pb_r_median": 0.8613637662637181,
+   "deg_top50_sum": 7,
+   "logfc_r": 0.10253144906014951,
+   "partial_r": 0.009339440904396172
+  }
+ },
+ "lora_n40__raw": {
+  "fold": {
+   "comp_l1": [
+    0.36688193889597037,
+    0.0725188939203545,
+    5
+   ],
+   "precision": [
+    0.9872794117647059,
+    0.003973461802107911,
+    5
+   ],
+   "coverage": [
+    0.866877450980392,
+    0.023951531680828825,
+    5
+   ],
+   "pb_r_median": [
+    0.8906885726019695,
+    0.0028460485682006635,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.3647568921706022,
+   "precision": 0.9216260162601626,
+   "coverage": 0.817560975609756,
+   "pb_r_median": 0.8911901532892592,
+   "deg_top50_sum": 22,
+   "logfc_r": 0.47588059575028696,
+   "partial_r": 0.4020213698715577
+  },
+  "train": {
+   "selected_steps": [
+    3000,
+    3750,
+    7000,
+    4000,
+    4000
+   ],
+   "inner_val": 0.11155371371922748,
+   "pretrained_inner_val": 0.12256578737869858,
+   "n_trainable": 1296224,
+   "n_runs": 5
+  }
+ },
+ "zs__cal_n40": {
+  "fold": {
+   "comp_l1": [
+    0.4064691137322477,
+    0.07004105568178039,
+    5
+   ],
+   "precision": [
+    0.9893676470588236,
+    0.0014662312908900784,
+    5
+   ],
+   "coverage": [
+    0.6441274509803921,
+    0.03926332354791746,
+    5
+   ],
+   "pb_r_median": [
+    0.8609687389468764,
+    0.0040621492777142935,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.4041847565927502,
+   "precision": 0.8992682926829269,
+   "coverage": 0.432520325203252,
+   "pb_r_median": 0.8618309547119146,
+   "deg_top50_sum": 7,
+   "logfc_r": 0.11880840796993535,
+   "partial_r": 0.028867502383777652
+  }
+ },
+ "zs__cal": {
+  "fold": {
+   "comp_l1": [
+    0.40361387110895874,
+    0.07055413299620945,
+    5
+   ],
+   "precision": [
+    0.9880980392156863,
+    0.002754851806814452,
+    5
+   ],
+   "coverage": [
+    0.6426127450980392,
+    0.04313190558409921,
+    5
+   ],
+   "pb_r_median": [
+    0.8605092817493066,
+    0.0025026115331507896,
+    5
+   ]
+  },
+  "pooled": {
+   "comp_l1": 0.40065324062398555,
+   "precision": 0.9034146341463415,
+   "coverage": 0.4308130081300813,
+   "pb_r_median": 0.8611254347365214,
+   "deg_top50_sum": 11,
+   "logfc_r": 0.12267970089558217,
+   "partial_r": 0.03938214909552809
+  }
+ }
+}

paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibration-budget/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/rankcal_20261001.md

@@ -0,0 +1,34 @@
+# ye2026_sepsis: LoRA rank sweep and equal-budget calibration (2026-10-01)
+
+Existing arms: scores/ye2026_sepsis/fold<k>, scores/ye2026_sepsis/pooled_deg; new: scores/ye2026_sepsis/lora_rank_20261001, scores/ye2026_sepsis/cal_budget_20261001.
+
+## LoRA rank (alpha = r; comparator = r 16 at the same n)
+
+| n | r | comp L1 fold mean ± SD (folds) | coverage (fold mean) | pooled comp L1 | pooled precision / coverage | pooled logFC r (T/B/Mono/NK) | partial r | top-50 DEG | selected step; inner-val loss | pooled per-donor L1 vs comparator (Δ, better in, Wilcoxon p) |
+|---|---|---|---|---|---|---|---|---|---|---|
+| 5 | 4 | 0.427 ± 0.060 (5) | 0.780 | 0.426 | 0.907 / 0.742 | 0.27 | 0.23 | 2 | 2000–3500; 0.11272 | -0.002 (44% better, p 0.97) |
+| 5 | 16 | 0.430 ± 0.074 (5) | 0.785 | 0.428 | 0.911 / 0.767 | 0.29 | 0.25 | 5 | 2000–3000; 0.11266 | (comparator) |
+| 5 | 64 | 0.432 ± 0.099 (5) | 0.783 | 0.429 | 0.906 / 0.766 | 0.31 | 0.26 | 1 | 2000–2750; 0.11259 | +0.001 (50% better, p 0.66) |
+| 20 | 4 | 0.379 ± 0.075 (5) | 0.833 | 0.378 | 0.929 / 0.769 | 0.31 | 0.24 | 14 | 2750–4000; 0.11193 | -0.003 (56% better, p 0.58) |
+| 20 | 16 | 0.382 ± 0.079 (5) | 0.841 | 0.380 | 0.927 / 0.807 | 0.37 | 0.29 | 18 | 2500–4000; 0.11187 | (comparator) |
+| 20 | 64 | 0.378 ± 0.075 (5) | 0.840 | 0.376 | 0.911 / 0.808 | 0.43 | 0.36 | 18 | 2500–4000; 0.11182 | -0.004 (59% better, p 0.046) |
+| all | 4 | 0.365 ± 0.068 (5) | 0.855 | 0.363 | 0.923 / 0.789 | 0.48 | 0.42 | 18 | 3750–10500; 0.11159 | +0.002 (49% better, p 0.77) |
+| all | 16 | 0.363 ± 0.072 (5) | 0.865 | 0.362 | 0.926 / 0.829 | 0.53 | 0.47 | 23 | 3750–13000; 0.11141 | (comparator) |
+| all | 64 | 0.369 ± 0.056 (5) | 0.869 | 0.367 | 0.907 / 0.824 | 0.55 | 0.49 | 26 | 3750–19500; 0.11129 | +0.005 (50% better, p 0.66) |
+
+## Zero-shot paired calibration at the LoRA pair budget (comparator = zs__cal_n<n>)
+
+| n | arm | comp L1 fold mean ± SD (folds) | coverage (fold mean) | pooled comp L1 | pooled precision / coverage | pooled logFC r (T/B/Mono/NK) | partial r | top-50 DEG | selected step; inner-val loss | pooled per-donor L1 vs comparator (Δ, better in, Wilcoxon p) |
+|---|---|---|---|---|---|---|---|---|---|---|
+| 5 | zs__cal_n5 | 0.413 ± 0.079 (5) | 0.648 | 0.412 | 0.898 / 0.435 | 0.08 | -0.02 | 12 |  | (comparator) |
+| 5 | lora_n5__raw | 0.430 ± 0.074 (5) | 0.785 | 0.428 | 0.911 / 0.767 | 0.29 | 0.25 | 5 | 2000–3000; 0.11266 | +0.016 (50% better, p 0.68) |
+| 10 | zs__cal_n10 | 0.415 ± 0.078 (5) | 0.646 | 0.414 | 0.896 / 0.441 | 0.10 | 0.00 | 10 |  | (comparator) |
+| 10 | lora_n10__raw | 0.392 ± 0.060 (5) | 0.824 | 0.390 | 0.916 / 0.786 | 0.42 | 0.37 | 8 | 2000–3250; 0.11227 | -0.024 (55% better, p 0.32) |
+| 20 | zs__cal_n20 | 0.409 ± 0.077 (5) | 0.637 | 0.407 | 0.899 / 0.425 | 0.10 | 0.01 | 7 |  | (comparator) |
+| 20 | lora_n20__raw | 0.382 ± 0.079 (5) | 0.841 | 0.380 | 0.927 / 0.807 | 0.37 | 0.29 | 18 | 2500–4000; 0.11187 | -0.027 (59% better, p 0.23) |
+| 40 | zs__cal_n40 | 0.406 ± 0.070 (5) | 0.644 | 0.404 | 0.899 / 0.433 | 0.12 | 0.03 | 7 |  | (comparator) |
+| 40 | lora_n40__raw | 0.367 ± 0.073 (5) | 0.867 | 0.365 | 0.922 / 0.818 | 0.48 | 0.40 | 22 | 3000–7000; 0.11155 | -0.039 (61% better, p 0.081) |
+| all | zs__cal | 0.404 ± 0.071 (5) | 0.643 | 0.401 | 0.903 / 0.431 | 0.12 | 0.04 | 11 |  | (comparator) |
+| all | real_lora__raw | 0.363 ± 0.072 (5) | 0.865 | 0.362 | 0.926 / 0.829 | 0.53 | 0.47 | 23 | 3750–13000; 0.11141 | -0.039 (54% better, p 0.19) |
+
+Fold scores use each fold's test donors as reference; pooled scores use all donors (cross-fitted). n = training pairs; every run also saw the fold's inner-validation pairs, which zs__cal_n<n> uses for alpha.