Create README.md

README.md

@@ -0,0 +1,221 @@
+# OCR-VQA Images — Download, Reassemble, and Place into LibMoE
+
+This guide shows how to download split archives from Hugging Face, merge them into a single ZIP, extract, and place the images into the LibMoE data tree:
+
+```
+libmoe/
+└── data/
+    ├── image_onevision/
+    ├── coco/
+    │   └── train2017/
+    ├── gqa/
+    │   └── images/
+    ├── ocr_vqa/
+    │   └── images/
+    ├── textvqa/
+    │   └── train_images/
+    └── vg/
+        ├── VG_100K/
+        └── VG_100K_2/
+```
+
+> Tested on Linux (bash). Requires ~100 GB free space for download + extraction room.
+
+---
+
+## 1) Prerequisites
+
+* Python + `huggingface_hub` CLI (or use `curl/wget`)
+* (Optional but faster) `hf_transfer` for high-speed downloads
+
+Install CLIs:
+
+```bash
+pip install -U huggingface_hub hf_transfer
+# Login once (stores token):
+huggingface-cli login   # or set HF_TOKEN env var
+```
+
+---
+
+## 2) Download split parts from Hugging Face
+
+Dataset: `DavidNguyen/ocr_vqa`, path: `ocr_vqa/images_part_*.zip.part`
+
+###  Using `curl` directly (works anywhere)
+
+```bash
+DEST=/cm/archive/namnv78_A100_PDM/data
+mkdir -p "$DEST/ocr_vqa_parts"
+cd "$DEST/ocr_vqa_parts"
+
+for p in {aa..ax}; do
+  # Direct "resolve" URLs on HF (replace if repo path changes)
+  curl -L -o "images_part_${p}.zip.part" \
+    "https://huggingface.co/datasets/DavidNguyen/ocr_vqa/resolve/main/ocr_vqa/images_part_${p}.zip.part?download=true"
+done
+
+ls -lh
+```
+
+---
+
+## 3) (Optional) Quick integrity check on parts
+
+```bash
+# Show sizes to spot any truncated files
+ls -lh images_part_*.zip.part
+
+# If you have reference checksums, verify:
+# md5sum -c md5sum_parts.txt
+```
+
+All parts should be ~2.1 GB except the last (`ax`) which may be smaller.
+
+---
+
+## 4) Concatenate parts → a single ZIP
+
+If you already downloaded to `/cm/archive/namnv78_A100_PDM/data/`, the one-liner below (as requested) works:
+
+```bash
+cat /cm/archive/namnv78_A100_PDM/data/images_part_*.zip.part \
+  > /cm/archive/namnv78_A100_PDM/data/images.zip
+```
+
+If you used the `ocr_vqa_parts` subfolder above:
+
+```bash
+cd /cm/archive/namnv78_A100_PDM/data/ocr_vqa_parts
+cat images_part_*.zip.part > ../images.zip
+```
+
+Verify the merged file:
+
+```bash
+cd /cm/archive/namnv78_A100_PDM/data
+ls -lh images.zip   # expect ~48–50 GB total
+# Optional checksum if provided:
+# md5sum images.zip
+```
+
+---
+
+## 5) Unzip
+
+```bash
+cd /cm/archive/namnv78_A100_PDM/data
+mkdir -p ./images_unzip
+unzip -q images.zip -d ./images_unzip
+# Or faster with 7z:
+# 7z x images.zip -o./images_unzip
+```
+
+After extraction you should see the image files or an `images/` folder inside `images_unzip/` (depends on how the ZIP was packed).
+
+---
+
+## 6) Place into the LibMoE data tree
+
+Target layout:
+
+```
+libmoe/data/ocr_vqa/images/
+```
+
+Create the tree (and optionally the other dataset folders to keep structure consistent):
+
+```bash
+# Adjust LIBMOE_ROOT to your actual path
+LIBMOE_ROOT=~/projects/libmoe
+
+mkdir -p "$LIBMOE_ROOT/data/image_onevision"
+mkdir -p "$LIBMOE_ROOT/data/coco/train2017"
+mkdir -p "$LIBMOE_ROOT/data/gqa/images"
+mkdir -p "$LIBMOE_ROOT/data/ocr_vqa/images"
+mkdir -p "$LIBMOE_ROOT/data/textvqa/train_images"
+mkdir -p "$LIBMOE_ROOT/data/vg/VG_100K"
+mkdir -p "$LIBMOE_ROOT/data/vg/VG_100K_2"
+```
+
+Now move (or symlink) the OCR-VQA images:
+
+```bash
+# Case A: the ZIP extracted directly to many image files:
+mv /cm/archive/namnv78_A100_PDM/data/images_unzip/* \
+   "$LIBMOE_ROOT/data/ocr_vqa/images/"
+
+# Case B: the ZIP contains an inner 'images/' folder:
+# mv /cm/archive/namnv78_A100_PDM/data/images_unzip/images/* \
+#    "$LIBMOE_ROOT/data/ocr_vqa/images/"
+```
+
+(If you prefer not to duplicate storage, create a symlink instead of `mv`):
+
+```bash
+# Remove the target if it exists and point a symlink:
+rm -rf "$LIBMOE_ROOT/data/ocr_vqa/images"
+ln -s /cm/archive/namnv78_A100_PDM/data/images_unzip \
+      "$LIBMOE_ROOT/data/ocr_vqa/images"
+```
+
+Check:
+
+```bash
+tree -L 2 "$LIBMOE_ROOT/data" | sed -n '1,200p'
+```
+
+You should now see:
+
+```
+libmoe/
+└── data/
+    ├── image_onevision/
+    ├── coco/
+    │   └── train2017/
+    ├── gqa/
+    │   └── images/
+    ├── ocr_vqa/
+    │   └── images/   <-- OCR-VQA images here
+    ├── textvqa/
+    │   └── train_images/
+    └── vg/
+        ├── VG_100K/
+        └── VG_100K_2/
+```
+
+---
+
+## 7) Troubleshooting
+
+* **`unzip: End-of-central-directory signature not found`**
+  One or more `.zip.part` files are missing/corrupted. Re-download the missing part(s) and re-run `cat`.
+
+* **Slow downloads / frequent timeouts**
+  Enable accelerated transfer: `export HF_HUB_ENABLE_HF_TRANSFER=1` (requires `hf_transfer` installed).
+  You can also run multiple parallel shells to fetch parts.
+
+* **Disk space**
+  You need free space for: all parts (~48–50 GB) **and** the unzipped images (size varies).
+  Delete parts after verifying extraction to reclaim space.
+
+* **Permission denied**
+  Use a writable directory or prefix commands with `sudo` if appropriate. Check that `$LIBMOE_ROOT` is correct.
+
+---
+
+## 8) Clean up (optional)
+
+```bash
+rm -f /cm/archive/namnv78_A100_PDM/data/images.zip
+rm -rf /cm/archive/namnv78_A100_PDM/data/ocr_vqa_parts
+```
+
+---
+
+## 9) Notes
+
+* If new parts (e.g., `ay`, `az`, `ba`, …) are added later, append them to the `PARTS=(...)` list before running the loop.
+* For reproducibility, consider publishing an `md5sum` file for all parts and the final ZIP.
+
+Happy hacking!