index: pinned historical training datasets and protected source-file snapshots

CHECKSUMS.sha256

@@ -0,0 +1,4 @@
+1051987a2cdeeeb668c457cf4818d46a1c8e2f5df40cbdcf98da68377dedf246  README.md
+2569cdcdb30128ffb7d1f178a2669580ead36e532f9f8c65a67f0f25c543caaf  TRAINING_DATA_INDEX.json
+67acb0f58a052ce7fd309e50cdaf473a9880bc31a8d1c7f68212d528088d275c  fetch_training_data.py
+4cd8ba59a93ab631419f29b61153ab8024cc7392f81fca1e0470c400144c8826  restore_legacy_mix6.py

README.md

@@ -0,0 +1,51 @@
+---
+license: other
+license_name: training-data-location-index-only
+license_link: https://huggingface.co/datasets/diffusion-gif/training-data-index/blob/main/README.md
+---
+# Previous training data: versioned location and checksum index
+
+`TRAINING_DATA_INDEX.json` tells you which retained training inputs belong to
+which model, their original/current snapshot roles, exact file SHA256/size,
+and an immutable owner-authorized private backup revision. The same index is
+attached at the roots of `diffusion-gif/any2ego` and `diffusion-gif/trick-prediction`.
+No original restricted annotation bodies, credentials or original machine paths
+are in this public index. This index is not a license grant for source data.
+
+## Important distinctions
+
+- Legacy mix6 SFT step900: train16,174 / heldout3,228; old multi-point protocol.
+- Legacy mix6 GRPO step250: the corresponding original train/heldout Parquet,
+  8 embedded observation-frame data URIs per record and identical SID sets.
+- All19,402 original mix6 NPZ observations and old JSON/Parquet are privately
+  preserved byte-for-byte, with independent tar-member checksum readback.
+- The later joint dataset31,202/5,015 and public simulation subset17,421 records
+  must not be substituted as the original training data of those old weights.
+- Any2Ego original JSON/CSV, full PhiSeg source manifests and the common32,558
+  export are separate snapshots. Captioner trainer args identify their source
+  JSON files; generator per-component launch-time subset hashes were not saved.
+
+The index records **retained file snapshot hashes**, not nonexistent hashes
+from the original training launch. The large Ego-Exo original RGB/GT-mask pool
+is NOT remotely backed up here. Original-run data identity and permanent
+third-party availability cannot be guaranteed merely by a JSON index.
+
+## Fetch your original training files
+
+Use an account authorized for the PRIVATE vault. Log in locally; do not put a
+real token in scripts or chat. This keeps gated/restricted annotations private.
+
+```bash
+pip install huggingface_hub
+hf auth login
+python fetch_training_data.py --index TRAINING_DATA_INDEX.json --dataset legacy_mix6_sft --output data_download
+# inspect the dry-run plan, then:
+python fetch_training_data.py --index TRAINING_DATA_INDEX.json --dataset legacy_mix6_sft --output data_download --execute
+python restore_legacy_mix6.py --snapshot-root data_download/snapshots/20261104 --output ready_mix6 --execute
+# GRPO: choose --dataset legacy_mix6_grpo; --dataset all_private fetches the entire protected snapshot.
+```
+
+Downloads are pinned to immutable commits, verify exact hashes, and never
+overwrite a different existing user file. The restore helper changes only
+`npz`/`frames_root` in WORKING JSON copies; all frozen originals remain intact.
+Keep a separate local/private copy for long-term availability.

fetch_training_data.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+"""Fetch pinned training files from the JSON index; dry-run unless --execute."""
+import argparse,hashlib,json,shutil
+from pathlib import Path,PurePosixPath
+
+
+def safe(value):
+ p=PurePosixPath(value)
+ if p.is_absolute() or '..' in p.parts or not p.parts:raise ValueError('Unsafe relative data path')
+ return Path(*p.parts)
+
+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 main():
+ p=argparse.ArgumentParser(description=__doc__)
+ directory=Path(__file__).absolute().parent
+ default_index=directory/'TRAINING_DATA_INDEX.json'
+ if not default_index.exists():default_index=directory.parent/'TRAINING_DATA_INDEX.json'
+ p.add_argument('--index',type=Path,default=default_index)
+ p.add_argument('--dataset',default='legacy_mix6_sft')
+ p.add_argument('--output',type=Path,default=Path('training_data_download'))
+ p.add_argument('--execute',action='store_true');a=p.parse_args()
+ index=json.loads(a.index.read_text());datasets={d['dataset_id']:d for d in index['datasets']}
+ if a.dataset=='all_private':ids={r['artifact_id'] for r in index['private_artifacts']}
+ elif a.dataset in datasets:ids=set(datasets[a.dataset].get('artifact_ids',[]))
+ else:raise ValueError('Unknown dataset ID; inspect index.datasets')
+ rows=[r for r in index['private_artifacts'] if r['artifact_id'] in ids]
+ if not rows:raise ValueError('No private snapshot artifacts for this dataset; inspect public download location')
+ source=index['private_snapshot']
+ if not source['repo_id'].startswith('diffusion-gif/training-data-vault-') or source['repo_type']!='dataset' or not source['private']:raise ValueError('Unexpected private data provider')
+ print(json.dumps({'dataset':a.dataset,'files':len(rows),'bytes':sum(r['bytes'] for r in rows),'source_repo':source['repo_id'],'revision':source['revision'],'login_required':True,'execute':a.execute},indent=2))
+ if not a.execute:return
+ from huggingface_hub import get_token,hf_hub_download
+ token=get_token()
+ if not token:raise RuntimeError('Run hf auth login locally using an account authorized for the private vault. Never put tokens in this file or chat.')
+ a.output.mkdir(parents=True,exist_ok=True)
+ for r in rows:
+  target=a.output/safe(r['path'])
+  if target.exists():
+   if not target.is_file() or target.stat().st_size!=r['bytes'] or sha(target)!=r['sha256']:raise ValueError('Existing user file differs; will not overwrite')
+   continue
+  cached=Path(hf_hub_download(repo_id=source['repo_id'],repo_type='dataset',revision=source['revision'],filename=r['path'],token=token,endpoint='https://huggingface.co'))
+  if cached.stat().st_size!=r['bytes'] or sha(cached)!=r['sha256']:raise ValueError('Downloaded file SHA256 mismatch')
+  target.parent.mkdir(parents=True,exist_ok=True);temporary=target.with_name(target.name+'.copy-partial')
+  shutil.copyfile(cached,temporary)
+  if sha(temporary)!=r['sha256']:raise ValueError('Copied file SHA256 mismatch')
+  temporary.replace(target);print('VERIFIED',r['path'],flush=True)
+ print('Done: all selected frozen files downloaded and SHA256 verified. See restore_legacy_mix6.py for a relocatable SFT working copy.')
+if __name__=='__main__':main()

restore_legacy_mix6.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""Restore legacy SFT observations/JSON working copies; never alter frozen files."""
+import argparse,hashlib,json,tarfile
+from pathlib import Path
+
+def main():
+ p=argparse.ArgumentParser(description=__doc__);p.add_argument('--snapshot-root',required=True,type=Path);p.add_argument('--output',required=True,type=Path);p.add_argument('--execute',action='store_true');a=p.parse_args()
+ src=a.snapshot_root/'legacy_mix6';locations={}
+ for line in (src/'observations/locations.private.jsonl').open():
+  r=json.loads(line);key=(r['split'],r['sid'])
+  if key in locations:raise ValueError('Duplicate observation ID')
+  if Path(r['archive']).name!=r['archive'] or Path(r['member']).name!=r['member']:raise ValueError('Unsafe archive/member path')
+  locations[key]=r
+ if not a.execute:
+  print(json.dumps({'records':len(locations),'output':str(a.output),'source_snapshot_modified':False,'execute':False},indent=2));return
+ if a.output.exists() and any(a.output.iterdir()):raise ValueError('Output must be a new/empty directory; will not overwrite user files')
+ a.output.mkdir(parents=True,exist_ok=True);by_archive={}
+ for key,r in locations.items():by_archive.setdefault(r['archive'],{})[r['member']]=(key,r)
+ for name,entries in by_archive.items():
+  with tarfile.open(src/'observations'/name,'r:') as archive:
+   for member in archive:
+    entry=entries.pop(member.name,None)
+    if entry is None or not member.isfile():raise ValueError('Unexpected archive entry')
+    (split,sid),r=entry;data=archive.extractfile(member).read()
+    if len(data)!=r['bytes'] or hashlib.sha256(data).hexdigest()!=r['sha256']:raise ValueError('Original NPZ checksum differs')
+    filename=hashlib.sha256(sid.encode()).hexdigest()[:32]+'.npz'
+    target=a.output/'frames'/split/filename;target.parent.mkdir(parents=True,exist_ok=True)
+    if target.exists():raise ValueError('Duplicate target frame file')
+    target.write_bytes(data)
+  if entries:raise ValueError('Missing observation member')
+ for split in ('train','heldout'):
+  with (src/'sft'/(split+'.jsonl')).open() as inp,(a.output/(split+'.jsonl')).open('w') as out:
+   for line in inp:
+    r=json.loads(line)
+    if (split,r['sid']) not in locations:raise ValueError('Input record lacks backed-up observation')
+    r['frames_root']=str((a.output/'frames'/split).resolve())
+    r['npz']=hashlib.sha256(r['sid'].encode()).hexdigest()[:32]+'.npz'
+    out.write(json.dumps(r,ensure_ascii=False)+'\n')
+ print('Restored exact NPZ bytes and re-based SFT train/heldout JSONL. Only npz/frames_root fields in WORKING copies changed; frozen originals are intact.')
+if __name__=='__main__':main()