Blame for tools/export_simulation.py

#!/usr/bin/env python3
"""Export ONLY licensed simulation records from a prepared joint SFT snapshot.

Original files are read-only. HMAC key/ID mapping stay in a separate private
work directory. Runtime account/API metadata and unknown-source media never
enter the public allowlist. This is not a full original-dataset anonymizer.
"""
import argparse,collections,concurrent.futures,hashlib,hmac,io,json,os,secrets,tarfile,zipfile
from pathlib import Path
import numpy as np
from PIL import Image
from schema import SOURCES,BODIES,TEXT_RISK,validate
BUDGET=1024**3

def sha(path):
 h=hashlib.sha256()
 with Path(path).open('rb') as f:
  for b in iter(lambda:f.read(8*1024*1024),b''):h.update(b)
 return h.hexdigest()

def dump(path,data):path.write_text(json.dumps(data,ensure_ascii=False,indent=2)+'\n')

def frame_bundle(source):
 path=Path(source['frames_root'])/source['npz']
 before=path.stat();buffer=io.BytesIO();images=[];encoded_hashes=[]
 with np.load(path,allow_pickle=False) as arrays,zipfile.ZipFile(buffer,'w',compression=zipfile.ZIP_DEFLATED,compresslevel=6) as output:
  if set(arrays.files)!={f'f{i}' for i in range(8)}:raise ValueError('Unknown source NPZ fields')
  for i in range(8):
   arr=arrays[f'f{i}']
   if arr.dtype!=np.uint8 or arr.ndim!=1 or not 100<=arr.size<=8*1024*1024:raise ValueError('Unexpected frame encoding')
   encoded=arr.tobytes();image=Image.open(io.BytesIO(encoded))
   if image.format not in ('JPEG','PNG') or image.getexif():raise ValueError('Unreviewed frame metadata/EXIF')
   if set(image.info)-{'jfif','jfif_version','jfif_unit','jfif_density','gamma','srgb'}:raise ValueError('Additional image metadata found')
   image.load()
   if list(reversed(image.size))!=source['hw']:raise ValueError('Source image dimensions differ')
   # Only pre-existing encoded observation bytes; no resize/re-encode or future frames.
   encoded_hashes.append(hashlib.sha256(encoded).hexdigest())
   if source.get('_make_preview'):images.append(image.convert('RGB').copy())
   npy=io.BytesIO();np.lib.format.write_array(npy,arr,allow_pickle=False)
   info=zipfile.ZipInfo(f'f{i}.npy',date_time=(1980,1,1,0,0,0));info.compress_type=zipfile.ZIP_DEFLATED;info.external_attr=0o644<<16
   output.writestr(info,npy.getvalue())
 after=path.stat()
 if (before.st_size,before.st_mtime_ns)!=(after.st_size,after.st_mtime_ns):raise ValueError('Source changed during read')
 return buffer.getvalue(),encoded_hashes,images

class Sharder:
 def __init__(self,root,split):self.root=root;self.split=split;self.index=-1;self.handle=None;self.rows=[];self.shards=[];self.size=0
 def close(self):
  if self.handle is None:return
  self.handle.close();temp=self.root/(self.name+'.partial');target=self.root/self.name
  entries={r['media_member']:r for r in self.rows}
  with tarfile.open(temp,'r:') as f:
   for member in f:
    if not member.isfile() or member.name not in entries:raise ValueError('Unexpected tar member')
    data=f.extractfile(member).read()
    if hashlib.sha256(data).hexdigest()!=entries.pop(member.name)['media_sha256']:raise ValueError('Tar member readback SHA256 differs')
  if entries:raise ValueError('Missing tar member')
  if target.exists():raise ValueError('Will not overwrite an existing archive')
  temp.replace(target);self.shards.append({'path':'data/archives/'+self.name,'bytes':target.stat().st_size,'sha256':sha(target),'members':len(self.rows)})
  self.handle=None;self.rows=[]
 def add(self,row,data):
  padded=((len(data)+511)//512)*512+512
  if self.handle is not None and self.size+padded>BUDGET:self.close()
  if self.handle is None:
   self.index+=1;self.name=f'{self.split}-{self.index:05d}.tar';self.size=0
   self.handle=tarfile.open(self.root/(self.name+'.partial'),'w:')
  row['media_archive']='data/archives/'+self.name
  info=tarfile.TarInfo(row['media_member']);info.size=len(data);info.mtime=0;info.uid=info.gid=0;info.uname=info.gname='';info.mode=0o644
  self.handle.addfile(info,io.BytesIO(data));self.size+=padded;self.rows.append(dict(row))


def main():
 p=argparse.ArgumentParser(description=__doc__);p.add_argument('--sft-root',required=True,type=Path);p.add_argument('--cot-run',required=True,type=Path);p.add_argument('--output',required=True,type=Path);p.add_argument('--privacy-work-root',required=True,type=Path);p.add_argument('--workers',type=int,default=4);a=p.parse_args()
 out=a.output.resolve();private=a.privacy_work_root.resolve()
 if private==out or out in private.parents:raise ValueError('Private mapping/key must be outside public payload')
 if (out/'PUBLIC_RELEASE_MANIFEST.json').exists():raise ValueError('Completed output already exists; use validator instead of overwriting')
 (out/'data/archives').mkdir(parents=True,exist_ok=True);(out/'previews').mkdir(exist_ok=True);private.mkdir(parents=True,exist_ok=True);private.chmod(0o700)
 if list((out/'data/archives').iterdir()):raise ValueError('Will not overwrite previous archives')
 key_path=private/'pseudonym_key'
 if not key_path.exists():key_path.write_bytes(secrets.token_bytes(32));key_path.chmod(0o600)
 key=key_path.read_bytes()
 def pseudonym(value):return hmac.new(key,value.encode(),hashlib.sha256).hexdigest()[:32]
 rows_by_split={};by_id={};groups={};excluded=collections.Counter();previews=set();source_counts={};media_rows=[];shards=[];seen=set()
 for split in ('train','heldout'):
  selected=[];group_set=set();counts=collections.Counter()
  for line in (a.sft_root/(split+'.jsonl')).open():
   s=json.loads(line);body=s.get('source')
   if body not in BODIES:excluded[str(body)]+=1;continue
   if s['joint_id'] in by_id:raise ValueError('Repeated original source ID')
   source='DexMimicGen' if body=='dmg' else 'RoboTwin2.0'
   row={k:s[k] for k in ('task','coordinate_space','point_names','n_hist','n_fut','forecast_arm_count','history','anchor','future_delta','hw')}
   row.update(sample_id=pseudonym('sample:'+s['joint_id']),group_id=pseudonym('group:'+body+':'+s['joint_episode_group']),split=split,source_dataset=source,license=SOURCES[source],robot_body=body,media_archive=f'data/archives/{split}-00000.tar',media_member='',media_sha256='0'*64)
   row['media_member']=row['sample_id']+'.npz';validate(row)
   if row['sample_id'] in seen:raise ValueError('Duplicate pseudonymous ID')
   seen.add(row['sample_id']);group_set.add(row['group_id']);counts[body]+=1
   s['_make_preview']=split=='train' and body not in previews
   if s['_make_preview']:previews.add(body)
   selected.append((s,row));by_id[s['joint_id']]=row
  rows_by_split[split]=selected;groups[split]=group_set;source_counts[split]=dict(counts)
 if groups['train']&groups['heldout']:raise ValueError('Source episode groups leak across splits')
 if (len(rows_by_split['train']),len(rows_by_split['heldout']))!=(14253,3168):raise ValueError('Unexpected reviewed simulation snapshot count')
 mapping=private/'id_mapping.jsonl'
 with mapping.open('w') as f:
  for original,row in by_id.items():f.write(json.dumps({'original_id':original,'public_id':row['sample_id']})+'\n')
 mapping.chmod(0o600)
 preview_rows=[]
 for split,selected in rows_by_split.items():
  sharder=Sharder(out/'data/archives',split)
  with (out/'data'/(split+'.jsonl')).open('w') as labels,concurrent.futures.ThreadPoolExecutor(max_workers=a.workers) as pool:
   for number,((source,row),(data,encoded_hashes,images)) in enumerate(zip(selected,pool.map(frame_bundle,(s for s,r in selected))),1):
    row['media_sha256']=hashlib.sha256(data).hexdigest();sharder.add(row,data);validate(row)
    labels.write(json.dumps(row,separators=(',',':'))+'\n')
    media_rows.append({'sample_id':row['sample_id'],'media_archive':row['media_archive'],'media_member':row['media_member'],'bytes':len(data),'sha256':row['media_sha256'],'encoded_frame_SHA256':encoded_hashes})
    if images:
     name=f'preview_{len(preview_rows)+1:02d}.gif';images[0].save(out/'previews'/name,save_all=True,append_images=images[1:],duration=180,loop=0)
     preview_rows.append({'path':'previews/'+name,'sample_id':row['sample_id'],'source_dataset':row['source_dataset'],'robot_body':row['robot_body'],'task':row['task'],'kind':'observed_simulation_frames_not_predictions'})
    if number%1000==0:print(f'{split}: {number}/{len(selected)} observations packaged and image metadata checked',flush=True)
  sharder.close();shards+=sharder.shards
 dump(out/'SHARDS.json',{'shards':shards,'every_tar_member_independently_SHA256_verified':True})
 with (out/'MEDIA_MANIFEST.jsonl').open('w') as f:
  for row in media_rows:f.write(json.dumps(row,separators=(',',':'))+'\n')
 dump(out/'previews/manifest.json',preview_rows)
 # Pick the exact publicly returned response version referenced by saved quality checks.
 reviews={}
 for line in (a.cot_run/'quality_reviews.jsonl').open():
  r=json.loads(line);reviews[r['id']]=r
 candidates={}
 for line in (a.cot_run/'candidates.jsonl').open():
  c=json.loads(line);r=reviews.get(c['id'])
  if c['id'] in by_id and r and hashlib.sha256(c['public_content'].encode()).hexdigest()==r['public_content_sha256']:candidates[c['id']]=c
 rationale_counts=collections.Counter();text_withheld=0
 with (out/'data/public_rationales.jsonl').open('w') as f:
  for original,c in candidates.items():
   text=c.get('reasoning') or '';row=by_id[original];review=reviews[original]
   if c.get('future_GT_sent_to_generator') is not False or not text or len(text)>20000 or TEXT_RISK.search(text) or original in text:
    text_withheld+=1;continue
   status='eligible_weak_draft' if review.get('eligible_for_weak_supervision') and c.get('machine_format_role_check_pass') else 'held_for_review'
   pub={'sample_id':row['sample_id'],'split':row['split'],'public_rationale':text,'status':status,'generation_source':'DeepSeek published API response','human_approved':False,'faithfulness_verified':False,'future_GT_sent_to_generator':False,'teacher_numeric_prediction_generated':False}
   f.write(json.dumps(pub,separators=(',',':'))+'\n');rationale_counts[status]+=1
 report={'scope':'licensed RoboTwin2.0 and DexMimicGen prepared simulation observation/trajectory subset',
         'counts':{k:len(v) for k,v in rows_by_split.items()},'by_body':source_counts,
         'raw_media_records':len(media_rows),'public_rationales':sum(rationale_counts.values()),'rationale_quality_counts':dict(rationale_counts),
         'rationales_withheld_for_sensitive_or_unverified_text':text_withheld,'excluded_unknown_license_sources':dict(excluded),
         'source_observation_image_bytes_preserved':True,'all_image_frames_decoded_and_metadata_checked':True,
         'image_EXIF_and_nonallowlisted_metadata_present':False,'original_IDs_paths_timestamps_account_API_metadata_exported':False,
         'key_or_mapping_exported':False,'each_tar_and_member_readback_SHA256_verified':True,
         'source_episode_groups_separate':True,'source_files_modified':False,'teacher_API_calls':0,'GPU_used':False,
         'full_original_356GB_collection_published':False,'new_training_or_model_predictions':False}
 dump(out/'SANITIZATION_REPORT.json',report)
 print(json.dumps(report,ensure_ascii=False,indent=2),flush=True)
if __name__=='__main__':main()