Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261006T033847Z-bulkonly-lora5-k4000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_k4000_20261006/analysis/benchmark/ye2026_sepsis/AGGREGATE.md

@@ -0,0 +1,26 @@
+# ye2026_sepsis: LoRA-5 at K = 4,000 vs K = 1,000 on the paired benchmark (5 draws)
+
+Same five few-pair draws (same 5 training donors each), same test donors, same scorer (paired_finetune_intradataset_20260927/tools/score_ft.py), 1,500 cells per test sample. Mean ± SD over draws.
+
+| metric | LoRA-5, K = 1,000 (benchmark) | LoRA-5, K = 4,000 | K4000 − K1000 (draws better / n) |
+|---|---:|---:|---|
+| comp_l1 | 0.466 ± 0.044 | 0.445 ± 0.051 | -0.022 (4 / 5) |
+| precision | 0.906 ± 0.020 | 0.910 ± 0.019 | +0.004 (2 / 5) |
+| coverage | 0.694 ± 0.042 | 0.698 ± 0.035 | +0.004 (3 / 5) |
+| coverage_dm | 0.834 ± 0.023 | 0.836 ± 0.023 | +0.001 (2 / 5) |
+| pb_r_median | 0.883 ± 0.003 | 0.884 ± 0.003 | +0.001 (3 / 5) |
+| mono_pct_gen | 17.839 ± 9.673 | 18.706 ± 10.052 | +0.867 (2 / 5) |
+| mono_pct_obs | 20.689 ± 0.444 | 20.689 ± 0.444 | +0.000 |
+| median_umi_gen | 1728.600 ± 189.579 | 1781.200 ± 169.896 | +52.600 (2 / 5) |
+| median_umi_obs | 1971.000 ± 18.841 | 1971.000 ± 18.841 | +0.000 |
+| deg_top50_sum | 16.400 ± 6.731 | 17.800 ± 9.445 | +1.400 (2 / 5) |
+| deg_top50_sum_broadcast | 20.000 ± 1.871 | 20.000 ± 1.871 | +0.000 |
+| logfc_r_gen | 0.350 ± 0.162 | 0.327 ± 0.128 | -0.023 (2 / 5) |
+| logfc_r_broadcast | 0.320 ± 0.024 | 0.320 ± 0.024 | +0.000 |
+| logfc_partial_r | 0.305 ± 0.189 | 0.276 ± 0.160 | -0.029 (2 / 5) |
+
+Per donor-draw composition L1, K = 4,000 − K = 1,000: -0.0215; K = 4,000 is better in 57% of 385 donor-draws (Wilcoxon p 0.0054).
+
+Lower is better for comp_l1; closer to *_obs is better for mono_pct_gen and median_umi_gen; higher is better for precision, coverage, pseudobulk r, the logFC r values and the top-50 DEG overlap. *_obs and *broadcast rows are the same for both arms (observed truth / bulk baseline).
+
+Scorer check: this plan's K = 1,000 arm vs the benchmark's stored draw scores, max |difference| per draw: draw0 0 (25 metrics), draw1 0 (25 metrics), draw2 0 (25 metrics), draw3 0 (25 metrics), draw4 0 (25 metrics).

paper-releases/pbisc-diffusion/updates/20261006T033847Z-bulkonly-lora5-k4000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_k4000_20261006/analysis/compare_k1000/imperial_eskd/SUMMARY.md

@@ -0,0 +1,42 @@
+# imperial_eskd: paired-test calibration, LoRA-5 K = 4,000 vs K = 1,000
+
+50 features x 5 draws = 250 pairs (same draws, test samples and observed truth). Level ok: ratio 0.5-2 and Spearman >= 0.3. Contrast ok: same sign and |generated difference| within 0.5-2x of observed.
+
+| draw | level ok K1000 | level ok K4000 | COVID+ vs COVID- ok K1000 | K4000 | severe/critical vs mild/moderate (COVID+) ok K1000 | K4000 |
+|---|---:|---:|---:|---:|---:|---:|
+| 0 | 21 | 23 | 19 | 18 | 18 | 20 |
+| 1 | 24 | 19 | 20 | 24 | 18 | 19 |
+| 2 | 25 | 25 | 19 | 20 | 20 | 18 |
+| 3 | 25 | 22 | 21 | 22 | 15 | 11 |
+| 4 | 16 | 21 | 20 | 22 | 15 | 18 |
+| all | 111 | 110 | 99 | 106 | 86 | 86 |
+
+- |log2 ratio| (distance of the generated level from observed), K4000 − K1000: median +0.008; closer in 48% of 250 pairs.
+- Spearman over test samples, K4000 − K1000: median +0.026; higher in 68% of 250 pairs.
+- COVID+ vs COVID-: same sign in 214 (K1000) vs 214 (K4000) of 250 pairs.
+- severe/critical vs mild/moderate (COVID+): same sign in 187 (K1000) vs 191 (K4000) of 250 pairs.
+
+Level gate gained at K4000 (10 features; draws): Cytotoxic CD4 T [1]; l1 | DC [4]; l1 | NK cell [0,2,3]; l2 | CD16 monocyte [4]; l2 | CD56dim NK cell [4]; l2 | Intermediate monocyte [1,2]; l2 | MAIT [0]; l2 | Naive B cell [2]; l2 | Naive CD8 T cell [4]; l2 | Proliferating T cell [4]
+
+Level gate lost at K4000 (13 features; draws): Cycling T/NK [2]; ISG-high T [1]; Plasmablast (label) [1]; Plasmablast (markers) [2]; l2 | CD16 monocyte [1]; l2 | Erythrocyte [2]; l2 | MAIT [3]; l2 | Memory B cell [1]; l2 | Plasma cell [1]; l2 | Platelet [3]; l2 | Progenitor cell [1]; l2 | Proliferating T cell [3]; l2 | pDC [1,3]
+
+Named features, ratio (Spearman) per draw:
+
+| feature | version | d0 | d1 | d2 | d3 | d4 |
+|---|---|---:|---:|---:|---:|---:|
+| l1 | Monocyte | k1000 | 1.08 (0.39) | 0.83 (0.33) | 1.11 (0.35) | 1.21 (0.42) | 0.98 (0.18) |
+| l1 | Monocyte | k4000 | 1.10 (0.45) | 0.91 (0.42) | 1.15 (0.38) | 1.30 (0.48) | 0.98 (0.27) |
+| l1 | NK cell | k1000 | 0.80 (0.20) | 0.43 (0.31) | 0.67 (0.21) | 1.09 (0.25) | 0.82 (0.14) |
+| l1 | NK cell | k4000 | 0.82 (0.31) | 0.40 (0.42) | 0.51 (0.32) | 1.03 (0.30) | 0.84 (0.25) |
+| l1 | DC | k1000 | 0.93 (0.23) | 1.08 (0.17) | 1.18 (0.13) | 0.59 (0.33) | 0.74 (0.28) |
+| l1 | DC | k4000 | 1.00 (0.23) | 1.15 (0.27) | 1.18 (0.21) | 0.66 (0.34) | 0.78 (0.33) |
+| l1 | B cell | k1000 | 0.54 (0.66) | 0.88 (0.54) | 0.82 (0.56) | 0.55 (0.62) | 1.17 (0.57) |
+| l1 | B cell | k4000 | 0.56 (0.67) | 1.00 (0.58) | 0.96 (0.59) | 0.56 (0.66) | 1.22 (0.62) |
+| l1 | T cell | k1000 | 1.13 (0.49) | 1.29 (0.41) | 1.10 (0.48) | 1.05 (0.45) | 1.03 (0.37) |
+| l1 | T cell | k4000 | 1.11 (0.53) | 1.24 (0.52) | 1.10 (0.51) | 1.03 (0.52) | 1.01 (0.49) |
+| l2 | MAIT | k1000 | 0.53 (0.19) | 0.41 (0.29) | 0.44 (0.39) | 0.54 (0.36) | 0.66 (0.15) |
+| l2 | MAIT | k4000 | 0.50 (0.34) | 0.40 (0.31) | 0.40 (0.44) | 0.61 (0.29) | 0.62 (0.15) |
+| l2 | CD14 monocyte | k1000 | 1.04 (0.48) | 1.05 (0.44) | 1.25 (0.42) | 1.21 (0.48) | 0.73 (0.33) |
+| l2 | CD14 monocyte | k4000 | 1.03 (0.50) | 1.24 (0.47) | 1.36 (0.42) | 1.31 (0.51) | 0.71 (0.33) |
+| l2 | Plasma cell | k1000 | 0.23 (0.28) | 0.66 (0.34) | 1.21 (0.20) | 0.61 (0.29) | 0.67 (0.43) |
+| l2 | Plasma cell | k4000 | 0.25 (0.28) | 0.63 (0.29) | 1.56 (0.22) | 0.53 (0.30) | 0.68 (0.46) |

paper-releases/pbisc-diffusion/updates/20261006T033847Z-bulkonly-lora5-k4000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_k4000_20261006/analysis/compare_k1000/ye2026_sepsis/SUMMARY.md

@@ -0,0 +1,43 @@
+# ye2026_sepsis: paired-test calibration, LoRA-5 K = 4,000 vs K = 1,000
+
+50 features x 5 draws = 250 pairs (same draws, test samples and observed truth). Level ok: ratio 0.5-2 and Spearman >= 0.3. Contrast ok: same sign and |generated difference| within 0.5-2x of observed.
+
+| draw | level ok K1000 | level ok K4000 | adult sepsis vs adult HC ok K1000 | K4000 | adult sepsis vs adult ICU control ok K1000 | K4000 | sepsis vs all controls ok K1000 | K4000 |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | 21 | 23 | 22 | 20 | 10 | 11 | 14 | 14 |
+| 1 | 18 | 16 | 11 | 13 | 9 | 6 | 12 | 9 |
+| 2 | 16 | 18 | 12 | 16 | 13 | 14 | 14 | 15 |
+| 3 | 22 | 27 | 12 | 20 | 8 | 16 | 13 | 12 |
+| 4 | 23 | 23 | 16 | 14 | 17 | 19 | 15 | 13 |
+| all | 100 | 107 | 73 | 83 | 57 | 66 | 68 | 63 |
+
+- |log2 ratio| (distance of the generated level from observed), K4000 − K1000: median -0.008; closer in 54% of 250 pairs.
+- Spearman over test samples, K4000 − K1000: median +0.013; higher in 60% of 250 pairs.
+- adult sepsis vs adult HC: same sign in 193 (K1000) vs 192 (K4000) of 250 pairs.
+- adult sepsis vs adult ICU control: same sign in 170 (K1000) vs 182 (K4000) of 250 pairs.
+- sepsis vs all controls: same sign in 165 (K1000) vs 169 (K4000) of 250 pairs.
+
+Level gate gained at K4000 (13 features; draws): ISG-high T [0]; Plasmablast (label) [3]; l1 | Other [2,3]; l1 | T cell [2,3]; l2 | CD56dim NK cell [1]; l2 | DN T cell [0,3]; l2 | Memory B cell [3]; l2 | Naive B cell [0]; l2 | Plasma cell [3]; l2 | Platelet [3]; l2 | Proliferating NK cell [0,2,4]; l2 | Treg [3]; l2 | pDC [3]
+
+Level gate lost at K4000 (11 features; draws): C1Q-high monocyte (reference) [4]; ISG-high T [1]; Proliferating T (label) [3]; l1 | Monocyte [0]; l1 | Other [1]; l2 | CD14 monocyte [0]; l2 | Effector B cell [2]; l2 | MAIT [3]; l2 | Platelet [1]; l2 | Proliferating T cell [3]; l2 | cDC2 [3]
+
+Named features, ratio (Spearman) per draw:
+
+| feature | version | d0 | d1 | d2 | d3 | d4 |
+|---|---|---:|---:|---:|---:|---:|
+| l1 | Monocyte | k1000 | 0.53 (0.45) | 1.70 (0.50) | 0.53 (0.53) | 0.71 (0.49) | 0.87 (0.50) |
+| l1 | Monocyte | k4000 | 0.45 (0.39) | 1.72 (0.50) | 0.53 (0.52) | 0.99 (0.54) | 0.86 (0.51) |
+| l1 | NK cell | k1000 | 0.87 (0.56) | 0.46 (0.43) | 0.46 (0.43) | 0.64 (0.46) | 0.52 (0.50) |
+| l1 | NK cell | k4000 | 0.91 (0.59) | 0.48 (0.50) | 0.49 (0.48) | 0.67 (0.51) | 0.52 (0.58) |
+| l1 | DC | k1000 | 0.67 (0.28) | 1.66 (0.20) | 0.43 (0.54) | 1.03 (0.56) | 0.87 (0.48) |
+| l1 | DC | k4000 | 0.58 (0.22) | 1.73 (0.15) | 0.41 (0.53) | 1.19 (0.38) | 0.82 (0.48) |
+| l1 | B cell | k1000 | 1.20 (0.83) | 0.56 (0.59) | 0.98 (0.75) | 0.55 (0.59) | 1.00 (0.82) |
+| l1 | B cell | k4000 | 1.15 (0.85) | 0.59 (0.66) | 0.94 (0.78) | 0.73 (0.81) | 0.99 (0.83) |
+| l1 | T cell | k1000 | 1.20 (0.60) | 0.92 (0.65) | 1.35 (0.29) | 1.37 (0.22) | 1.16 (0.49) |
+| l1 | T cell | k4000 | 1.25 (0.56) | 0.92 (0.66) | 1.35 (0.31) | 1.17 (0.62) | 1.16 (0.60) |
+| l2 | MAIT | k1000 | 0.53 (0.46) | 0.52 (0.36) | 0.55 (0.43) | 0.55 (0.38) | 0.55 (0.45) |
+| l2 | MAIT | k4000 | 0.54 (0.52) | 0.57 (0.36) | 0.53 (0.47) | 0.49 (0.40) | 0.58 (0.51) |
+| l2 | CD14 monocyte | k1000 | 0.53 (0.33) | 1.08 (0.46) | 0.51 (0.53) | 0.67 (0.48) | 0.76 (0.46) |
+| l2 | CD14 monocyte | k4000 | 0.45 (0.27) | 1.20 (0.47) | 0.52 (0.54) | 1.01 (0.52) | 0.76 (0.51) |
+| l2 | Plasma cell | k1000 | 0.38 (0.39) | 0.25 (0.22) | 0.15 (0.21) | 0.46 (0.26) | 0.43 (0.33) |
+| l2 | Plasma cell | k4000 | 0.38 (0.40) | 0.25 (0.25) | 0.18 (0.19) | 0.78 (0.34) | 0.45 (0.31) |

paper-releases/pbisc-diffusion/updates/20261006T033847Z-bulkonly-lora5-k4000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/K_BUILD.json

@@ -0,0 +1,112 @@
+{
+ "dataset": "imperial_eskd",
+ "k": 4000,
+ "cells_per_bulk": {
+  "training": "unchanged (not used by ft_sdd.py)",
+  "paired": 1500,
+  "bulkonly": 4000
+ },
+ "written_utc": "2026-10-05T17:38:27.256835+00:00",
+ "tool": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/tools/build_k_conditions.py",
+ "tool_sha256": "ef86fa3ba086ca31ada2c85d930e216608b3199da5ff0463cc17c4471ddda92b",
+ "builder": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/realbulk_generation_20260922/build_realbulk_conditions.py",
+ "builder_sha256": "e818d69e6801b5699579db7fc450590a266708d1fc2d544743e8624c9ee6a30a",
+ "prepare_dataset_sha256": "36384262b23cf6260eff645812911c1a44f791ca8e4182765d91f2aabb3ccaef",
+ "first_plan_build_conditions_sha256": "38cb97593c9a916b0dcd8d470a8c756e65c7ccc4bee6193f688e509d2e922862",
+ "depth_anchor": 3869.603120208939,
+ "checks": {
+  "training_K1000_equals_conditions_raw_fixed": true,
+  "paired_draw0_K1000_equals_test_raw": true,
+  "paired_draw1_K1000_equals_test_raw": true,
+  "paired_draw2_K1000_equals_test_raw": true,
+  "paired_draw3_K1000_equals_test_raw": true,
+  "paired_draw4_K1000_equals_test_raw": true,
+  "bulkonly_K1000_equals_bulkonly_raw": true
+ },
+ "files": {
+  "training": {
+   "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/conditions_raw_fixed_K4000.npz",
+   "sha256": "d3039d2d90c19f419b797c57866ff707929f10af683fed64f13ef238d081aa0f",
+   "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/conditions_raw_fixed.npz",
+   "source_sha256": "41a62291e98e367d05108a0a95a859ef9ca1997f1fda4656cba58eeefbe56442",
+   "rows": 172,
+   "recipe_k": 4000,
+   "cells_per_bulk": [
+    1500
+   ]
+  },
+  "paired": {
+   "0": {
+    "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/paired_test/draw0_test_raw_K4000.npz",
+    "sha256": "e9e22b0aa245c11864057c6a2a27c0a622d39f09f16f8d44481976f3ff7be65e",
+    "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/draw0/test_raw.npz",
+    "source_sha256": "735672055b9222e45379a85fb0de4994028d33943ba220f048c7b6eafca65e0a",
+    "rows": 142,
+    "recipe_k": 4000,
+    "cells_per_bulk": [
+     1500
+    ]
+   },
+   "1": {
+    "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/paired_test/draw1_test_raw_K4000.npz",
+    "sha256": "9184925d5c7c6b5483d367fe42779684d1996c4720c19e5005a117123566ce31",
+    "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/draw1/test_raw.npz",
+    "source_sha256": "c2293bb938f0dbf9c76063b305230e4ad0b9bec482e4fe5dfcf38a110511283f",
+    "rows": 143,
+    "recipe_k": 4000,
+    "cells_per_bulk": [
+     1500
+    ]
+   },
+   "2": {
+    "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/paired_test/draw2_test_raw_K4000.npz",
+    "sha256": "5751bcdd56bdee7edef0b0978193d3899becf8a8f321e1198647082cc73c13fa",
+    "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/draw2/test_raw.npz",
+    "source_sha256": "11ba7c4e0f734b8eeab115536093d16205a1f592833e54115aa10cafd131e466",
+    "rows": 142,
+    "recipe_k": 4000,
+    "cells_per_bulk": [
+     1500
+    ]
+   },
+   "3": {
+    "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/paired_test/draw3_test_raw_K4000.npz",
+    "sha256": "375f75ac1f6094649e4cf0faf32c1b72bbaa82038879715b46410d8a21556921",
+    "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/draw3/test_raw.npz",
+    "source_sha256": "1bb5ea46702a37ba35cf81b4e82426c5cfa539ef881c477be37e40621df26b7f",
+    "rows": 145,
+    "recipe_k": 4000,
+    "cells_per_bulk": [
+     1500
+    ]
+   },
+   "4": {
+    "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/paired_test/draw4_test_raw_K4000.npz",
+    "sha256": "73869f7b529e5c2660089321efe2bebb373a0b0efbfcf9b78eaa3509b590d339",
+    "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/draw4/test_raw.npz",
+    "source_sha256": "55970c18927ba5d8e151bf797760d36017fffd360229507c346e39c3c6759f69",
+    "rows": 143,
+    "recipe_k": 4000,
+    "cells_per_bulk": [
+     1500
+    ]
+   }
+  },
+  "bulkonly": {
+   "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/bulkonly_raw_K4000.npz",
+   "sha256": "c6be369cae9e68f479af1c11eb83a8191b4a5c1a84518a755a993873c9c97434",
+   "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/data/imperial_eskd/bulkonly_raw.npz",
+   "source_sha256": "4ee0c1d4332cce6d0abd135c1d438510217f74a4ca007d0b7d8ba1350e5688f6",
+   "rows": 164,
+   "recipe_k": 4000,
+   "cells_per_bulk": [
+    4000
+   ]
+  },
+  "bulkonly_meta.csv": {
+   "file": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_k4000_20261006/data/imperial_eskd/bulkonly_meta.csv",
+   "sha256": "a03307389c482fb5aeb9ee439cf3ec485dfa92dc1ed7eac34350e483a84563f3",
+   "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/data/imperial_eskd/bulkonly_meta.csv"
+  }
+ }
+}