---
pretty_name: Prepared INSID3 benchmark datasets
---
# Prepared INSID3 benchmark datasets

ZIPs of extracted/processed benchmark folders, excluding original downloaded archives, embeddings, weights, logs and unfinished downloads. Each dataset is one ZIP64 archive compressed with DEFLATE level 1.

Extract ZIPs into the same `datasets` directory, then run `python datasets/DATASET/restore_layout.py` for each extracted dataset. This restores hardlinked aliases and relative symlinks without duplicating shared image bytes. PACO-Part requires LVIS for shared COCO images; remote_sensing contains the iSAID compatibility link. COCO2014 split definitions are materialized, so no server-specific external path is required.

Example:
```bash
unzip SUIM.zip -d datasets
python datasets/SUIM/restore_layout.py
```

`prepared_inventory.json` lists actual archive sizes and unique payload sizes. These are prepared folders, not the original raw dataset download archives.