Upload RESTORE.md with huggingface_hub

RESTORE.md

@@ -1,9 +1,11 @@
 # HiReT / Clinical-Reasoning backup — restore instructions
 
 ## What this is
-One AES-256 encrypted 7z archive, split into 8 volumes (40 GB each):
+One AES-256 encrypted 7z archive, split into **61 volumes of ~5 GB each**
+(40 GB volumes could not be uploaded reliably because the build container has a
+2 GiB memory cap):
 
-    hitraj_backup_20260929.7z.001 ... .008
+    hitraj_backup_20260929.7z.001 ... .061
 
 Archive totals: 444,062 files, 7,249 folders, 325.6 GB uncompressed.
 Everything was packed relative to `/root`, so extracting inside `/root`
@@ -13,31 +15,36 @@ restores the original absolute paths:
     /root/autodl-tmp/ClinicalReasoning-ICLR/...    (ICLR paper source and figures)
     /root/autodl-tmp/backup_meta/...               (env exports, system info)
 
+The passphrase is NOT stored in this archive or in the repository. Keep it in a
+password manager and in one offline copy; without it the backup cannot be opened.
+
 ## Restore
-1. Download ALL eight volumes plus this file and the checksum file into one folder.
+1. Download ALL 61 volumes plus this file and the checksum file into one folder.
 2. Verify volume integrity:
 
        sha256sum -c hitraj_backup_20260929.7z.volumes.sha256
 
-3. Extract (one command, all volumes are read automatically):
+3. Optional: rebuild the single-file archive exactly as it was created:
 
-       cd /root && 7z x /path/to/hitraj_backup_20260929.7z.001
+       cat hitraj_backup_20260929.7z.0* > hitraj_backup_20260929.7z
 
-   The passphrase is required. It is kept offline; it is NOT in this archive.
+4. Extract (one command; all volumes are read automatically):
+
+       cd /root && 7z x /path/to/hitraj_backup_20260929.7z.001
 
-4. Rebuild the Python environment (do not copy the old conda dir):
+5. Rebuild the Python environment (do not copy the old conda directory):
 
        /root/miniconda3/bin/conda env create -f /root/autodl-tmp/backup_meta/env_llama-fac-wxr.yml
        # or: pip install -r /root/autodl-tmp/backup_meta/pip_freeze.txt
 
-## Things that were NOT included (rebuildable or junk)
-- install_tmp/ and pip_cache/ (figure tooling and pip cache, 7.6 GB)
+## Not included (rebuildable or junk)
+- install_tmp/ and pip_cache/ (figure tooling and pip cache, ~7.6 GB)
 - __pycache__, *.lock, vscode-*.sock, torchinductor_root, .ipynb_checkpoints
 - /root/miniconda3 itself (rebuild from env_llama-fac-wxr.yml)
 
 ## Notes
-- The 6-class and LOS embedding caches ARE included (they take roughly 12-20 h of GPU
-  time to rebuild). Cache fingerprints live under each cache directory manifest.json.
-- Patient-level data (Datasets, per-sample predictions, caches) is MIMIC-IV derived:
-  keep the repository access restricted even though the archive is encrypted.
+- The 6-class and LOS embedding caches ARE included (rebuilding them costs roughly
+  12-20 h of GPU time). Cache fingerprints live in each cache directory manifest.json.
+- Datasets/, per-sample predictions and caches are MIMIC-IV derived: keep repository
+  access restricted and keep the passphrase private even though the archive is encrypted.
 - Historical configs use absolute paths; keep the /root/autodl-tmp layout.