Archive 4 checkpoints (1-4)

.gitattributes

@@ -260,3 +260,4 @@ paper-releases/pbisc-diffusion/updates/20261013T094815Z-all-celltypes-final-pale
 paper-releases/pbisc-diffusion/updates/20261013T094815Z-all-celltypes-final-palette-142/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/final_palette_20261013/native/gene_length_input.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261013T115045Z-figure-strengthening-154/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013_strengthened.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261013T115045Z-figure-strengthening-154/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013_strengthened.pptx filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261013T144247Z-f01-trunk4x-campaign-launch/diffusion/dataset/server_1/model8/plans/f01_trunk4x_complete_20261013/runtime/sshfs filter=lfs diff=lfs merge=lfs -text

paper-releases/pbisc-diffusion/updates/20261013T144247Z-f01-trunk4x-campaign-launch/diffusion/dataset/server_1/model8/plans/f01_trunk4x_complete_20261013/runtime/sshfs

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d42d5b5dc641c15ed40c8ea8148ff18d53d8e7c53b4034a366ce4ae4b2da30ef
+size 135736

paper-releases/pbisc-diffusion/updates/20261013T144247Z-f01-trunk4x-campaign-launch/diffusion/dataset/server_1/model8/plans/f01_trunk4x_complete_20261013/task_plan.md

@@ -0,0 +1,14 @@
+# Exact f01 6k trunk×4 campaign
+
+User request: pretrain trunk×4 with the paper's current×1 checkpoint's exact6k gene context, then prepare the paper comparisons; prefer server_2, fallback server_1 if busy.
+
+- [x] Inspect both hosts: both busy; select server_1 with exclusive-lock queue, preserve existing jobs.
+- [x] Freeze baseline config/data/code and prove only trunk architecture changes scientifically.
+- [x] Define full experiment matrix, dependencies, reuse and missing-evidence gates.
+- [x] Implement pretraining launcher and automatic continuation into matched downstream work.
+- [x] Run CPU contract checks and launch durable guarded queue.
+- [ ] Archive plan/code/launch evidence with truthful running/queued state.
+
+Not an early claim of completed training. Do not substitute the legacy-panel trunk checkpoint, shrink the gene panel, or rerun unchanged baselines.
+
+Current:128-step GPU profile passed; full pretraining started on server_1. Results pending. First launch stopped at environment gate before CUDA; logs preserved and explicit env corrected.

paper-releases/pbisc-diffusion/updates/20261013T144247Z-f01-trunk4x-campaign-launch/diffusion/dataset/server_1/model8/plans/f01_trunk4x_complete_20261013/tools/archive_completed.py

@@ -0,0 +1,47 @@
+"""Task-end paper archive, exact completed outputs only; large cell arrays as locators."""
+from pathlib import Path
+from datetime import datetime,timezone
+import json,subprocess,sys
+P=Path(__file__).resolve().parents[1];M=P.parents[1];RELEASE=M.parents[1]/'paper-release-20260928';CLI=RELEASE/'maintenance/update.py'
+result=json.loads((P/'RESULTS_STATUS.json').read_text());assert result['phase']=='data_complete_with_review_gates'
+assets=[];locations=[];seen=set();loc_seen=set()
+def add(path,kind):
+ p=path.resolve()
+ if str(p) not in seen:assets.append({'host':'server_1','source':str(p),'kind':kind,'state':'completed'});seen.add(str(p))
+parent=json.loads((P/'PRETRAIN_VERIFIED.json').read_text());add(Path(parent['checkpoint']),'checkpoint')
+for p in [P/'PRETRAIN_VERIFIED.json',P/'RESULTS_STATUS.json',P/'PENDING_EVIDENCE.json',P/'MATRIX.json',P/'CONTRACT.json']:
+ add(p,'report')
+for p in (P/'comparison').glob('*'):
+ if p.is_file():add(p,'figure' if p.suffix in ['.png','.pdf','.svg'] else 'table')
+for j in json.loads((P/'MATRIX.json').read_text())['paired_jobs']:
+ out=Path(j['output'])
+ if not j.get('adapter_alias'):
+  add(out/'run/checkpoint_ft.pt','checkpoint');add(out/'run/RUN.json','report')
+ for p in (out/'input').glob('**/ISOLATION.json'):add(p,'report')
+ for p in (out/'evaluation').glob('**/*.json'):
+  if '/labels/' not in str(p):add(p,'report')
+ for p in (out/'evaluation').glob('**/*.csv'):
+  if p.stat().st_size<4*2**20:add(p,'table')
+for p in (P/'payload').glob('**/*.npz'):
+ if p.with_suffix('.log').is_file():
+  r=json.loads(p.with_suffix('.log').read_text())
+  if r.get('output_sha256') and str(p.resolve()) not in loc_seen:
+   loc_seen.add(str(p.resolve()));locations.append({'host':'server_1','path':str(p.resolve()),'state':'completed','receipt':str(p.with_suffix('.log').resolve()),'sha256':r['output_sha256'],'reason':'Completed exact-f01-panel trunkx4 generation; preserved on verified NAS mount. Aliases are not independent experiments.'})
+# Keep all16 pretraining snapshots on NAS; publish the fixed E16 model and adaptation models.
+# Other epoch checkpoints are recorded, not silently deleted to reduce storage.
+pre=Path(parent['checkpoint']).parent
+snapshot={'epoch_checkpoints':[str(p) for p in sorted(pre.glob('checkpoint_epoch_*.pt'))],'retained_on_NAS':True,'selected_for_model_publication':parent['checkpoint']}
+(P/'PRETRAIN_SNAPSHOT_LOCATIONS.json').write_text(json.dumps(snapshot,indent=2)+'\n');add(P/'PRETRAIN_SNAPSHOT_LOCATIONS.json','report')
+request={'project_id':'pbisc-diffusion','summary':'Exact current f01 6k-panel trunkx4 pretraining and matched downstream data. Biological population identity/clinical reproduction gates remain explicitly pending where documented.','assets':assets,'inference_locations':locations}
+path=P/'FINAL_ARCHIVE_REQUEST.json';path.write_text(json.dumps(request,indent=2)+'\n');ident=datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%SZ')+'-f01-trunk4x-data'
+need=sum(Path(x['source']).stat().st_size for x in assets)
+free=int(subprocess.check_output(['ssh','spark-0110',"python3 -c 'import shutil;print(shutil.disk_usage(\"/home/sj_server_2/PBISC/paper-release-20260928\").free)'"],text=True))
+if free<need+32*2**30:
+ (P/'FINAL_ARCHIVE.json').write_text(json.dumps({'state':'pending_capacity','needed_bytes':need,'archive_free_bytes':free,'all_outputs_preserved_on_NAS':True},indent=2)+'\n');raise SystemExit('Archive capacity gate: no originals deleted; inspect storage routing')
+with (P/'final_archive.log').open('x') as log:
+ subprocess.run([sys.executable,str(CLI),'prepare','--request',str(path),'--id',ident],check=True,stdout=log,stderr=subprocess.STDOUT)
+ manifest=json.loads((RELEASE/'maintenance/updates'/ident/'manifest.json').read_text());assert {(x['host'],x['source']) for x in manifest['assets'] if x['category']!='update_metadata'} <= {(x['host'],x['source']) for x in assets}
+ subprocess.run([sys.executable,str(CLI),'publish',ident,'--execute'],check=True,stdout=log,stderr=subprocess.STDOUT)
+receipt=RELEASE/'maintenance/updates'/ident/'UPDATE_VERIFIED.json';assert json.loads(receipt.read_text())['status']=='PASS'
+assert json.loads((RELEASE/'LATEST.json').read_text())['id']==ident and json.loads((RELEASE/'CURRENT_CATALOG.json').read_text())['latest_id']==ident
+(P/'FINAL_ARCHIVE.json').write_text(json.dumps({'state':'verified','id':ident,'receipt':str(receipt),'data_readiness':result['phase']},indent=2)+'\n')

paper-releases/pbisc-diffusion/updates/20261013T144247Z-f01-trunk4x-campaign-launch/diffusion/dataset/server_1/model8/plans/f01_trunk4x_complete_20261013/tools/build_matrix.py

@@ -0,0 +1,49 @@
+"""Freeze paired adaptation matrix from actual current-paper run records."""
+from pathlib import Path
+import json,hashlib,subprocess,shlex
+import pandas as pd
+P=Path(__file__).resolve().parents[1];M=P.parents[1];F=M/'plans/paired_finetune_intradataset_20260927';S=M/'plans/paper_figures_strictpairs_20261012'
+COHORTS=['immicro','ye2026_sepsis','imperial_eskd','giroux2022','ji2025_chili']
+def sha(p):
+ with Path(p).open('rb') as f:return hashlib.file_digest(f,'sha256').hexdigest()
+def read_run(path):
+ if path.is_file():raw=path.read_text();host='server_1'
+ else:raw=subprocess.check_output(['ssh','spark-0110','cat '+shlex.quote(str(path))],text=True);host='server_2'
+ r=json.loads(raw);assert r['source_checkpoint_sha256']=='8c7cd0e83e2bc19fe03dfcbc084919a1e2f61be8cc8bc12f4caf543dfda91633'
+ dest=P/'baseline_records'/path.relative_to(F/'runs');dest.parent.mkdir(parents=True,exist_ok=True);dest.write_text(raw)
+ return r,{'host':host,'path':str(path),'snapshot':str(dest),'sha256':sha(dest)}
+def main():
+ jobs=[];baseline_records=[]
+ for support_path in sorted((S/'protocol/splits').glob('**/support.json')):
+  s=json.loads(support_path.read_text());rel=str(support_path.parent.relative_to(S/'protocol/splits'));name='strict/'+rel
+  jobs.append({'id':name,'family':'strict','cohort':s['cohort'],'support_path':str(support_path),'support_sha256':s['sha256'],'train_samples':s['train_sample_ids'],'validation_samples':[],'test_samples':s['test_sample_ids'],'train_donors':s['train_donor_ids'],'test_donors':s['test_donor_ids'],'condition':str(F/f"data/{s['cohort']}/conditions_raw_fixed.npz"),'test_condition':str(S/'data'/rel/'test_raw.npz'),'cal_condition':str(S/'data'/rel/'test_cal.npz'),'mode':'lora','rank':16,'alpha':16.,'lr':.0005,'warmup':200,'batch_size':256,'ema_decay':.999,'fixed_steps':3000,'seed':20260927,'sampling_seed':20260928,'baseline_score':str(S/'scores'/rel/'score/summary.json'),'baseline_generated':str(S/'generated'/rel/'lora.npz'),'scsemi_generated':str(S/'generated'/rel/'scsemi.npz')})
+ for cohort in COHORTS:
+  for fold in range(5):
+   for arm in (['real_lora','lora_r4','lora_r64','real_full'] if cohort=='immicro' else ['real_lora']):
+    path=F/f'runs/{cohort}/fold{fold}/{arm}/RUN.json';r,source=read_run(path);baseline_records.append(source)
+    job={'id':f'cv/{cohort}/{arm}/fold{fold}','family':'cv','cohort':cohort,'fold':fold,'old_arm':arm,'source_run':source,'train_samples':r['train_rows'],'validation_samples':r['inner_val_rows'],'test_samples':r['test_rows_never_loaded_for_training'],'condition':r['condition_npz'],'test_condition':str(F/f'data/{cohort}/fold{fold}/test_raw.npz'),'mode':r['mode'],'rank':r['trainable'].get('lora_rank',16),'alpha':r['trainable'].get('lora_alpha',16.),'lr':r['lr'],'warmup':r['warmup'],'batch_size':r['batch_size'],'ema_decay':r['ema_decay'],'fixed_steps':0,'seed':r['seed'],'max_steps':r['max_steps'],'eval_every':r['eval_every'],'patience':r['patience'],'sampling_seed':20260927}
+    # Read actual sampling seed when a receipt is locally available; otherwise the historical queue contract applies.
+    for candidate in [F/f'generated/{cohort}/fold{fold}/{arm}__raw.log']:
+     if candidate.is_file():job['sampling_seed']=json.loads(candidate.read_text())['seed']
+    jobs.append(job)
+  for draw in range(5):
+   path=F/f'runs/{cohort}/fewshot_n5/draw{draw}/lora/RUN.json';r,source=read_run(path);baseline_records.append(source)
+   alias=None
+   for x in jobs:
+    if x['family']=='strict' and x['cohort']==cohort and set(x['train_samples'])==set(r['train_rows']) and set(x['test_samples'])==set(r['test_rows_never_loaded_for_training']) and not r['inner_val_rows'] and r['fixed_steps']==3000:alias=x['id'];break
+   folds=pd.read_csv(F/f'data/{cohort}/folds.csv');lookup=folds.set_index('sample_id').donor_id.to_dict()
+   jobs.append({'id':f'legacy5/{cohort}/draw{draw}','family':'legacy5','cohort':cohort,'draw':draw,'source_run':source,'adapter_alias':alias,'train_samples':r['train_rows'],'train_donors':sorted({lookup[x] for x in r['train_rows']}),'validation_samples':[],'test_samples':r['test_rows_never_loaded_for_training'],'condition':r['condition_npz'],'test_condition':str(F/f'data/{cohort}/fewshot_n5/draw{draw}/test_raw.npz'),'mode':'lora','rank':16,'alpha':16.,'lr':r['lr'],'warmup':r['warmup'],'batch_size':r['batch_size'],'ema_decay':r['ema_decay'],'fixed_steps':3000,'seed':r['seed'],'sampling_seed':20260928,'baseline_score':str(F/f'scores/{cohort}/fewshot_n5/draw{draw}/summary.json')})
+ for job in jobs:
+  assert not set(job['train_samples'])&set(job['test_samples'])
+  assert not set(job['validation_samples'])&set(job['test_samples'])
+  for k in ['condition','test_condition']:
+   assert Path(job[k]).is_file(),job[k]
+   job[k+'_sha256']=sha(job[k])
+  job['output']=str(P/'payload/paired'/job['id'])
+ slereceipts=[]
+ for i in range(5):
+  tag='f016k_e16' if i==0 else f'f016ks{i}_e16';p=M/f'analysis/floor01_benchmark_20260923/generated_{tag}.log';r=json.loads(p.read_text());slereceipts.append({'seed':r['seed'],'baseline_receipt':str(p),'baseline_generated':str(p.with_suffix('.npz')),'seed_index':i})
+ result={'schema':'pbisc.f01.trunk4x.matrix.v1','paired_jobs':jobs,'new_adapter_count':sum(not j.get('adapter_alias') for j in jobs),'adapter_alias_count':sum(bool(j.get('adapter_alias')) for j in jobs),'sle':slereceipts,'global_cohorts':COHORTS,'bulkonly_cohorts':['ye2026_sepsis','imperial_eskd'],'rare_identity_gate':'Original-paper fine labels/timing required. Marker diagnostics do not complete biological reproduction.','comparison_scope':'Exact current f01 6000genes, no intersection projection','baseline_run_snapshots':baseline_records}
+ (P/'MATRIX.json').write_text(json.dumps(result,indent=2)+'\n')
+ print(json.dumps({'jobs':len(jobs),'adapters':result['new_adapter_count'],'aliases':result['adapter_alias_count'],'families':{f:sum(j['family']==f for j in jobs) for f in ['strict','cv','legacy5']}}))
+if __name__=='__main__':main()