diff --git a/.gitattributes b/.gitattributes
index 48edf9969ff2af8c04445bf1e5e74f779606c169..57f1e7530704c75979ca042fe9cc3561ac4ee484 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -4852,3 +4852,45 @@ documents/10.1056_nejmoa2302312/sap.pdf filter=lfs diff=lfs merge=lfs -text
 documents/10.1056_nejmoa2116596/sap.pdf filter=lfs diff=lfs merge=lfs -text
 documents/10.1016_s0140-6736(24)01498-3/sap.pdf filter=lfs diff=lfs merge=lfs -text
 documents/10.1056_nejmoa2210645/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(17)33297-x/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(17)33297-x/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.1660/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.0071/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.1660/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(18)32321-3/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(18)32321-3/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.0071/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1908655/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1908655/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1805104/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1816714/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1805104/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1816714/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1200_jco.2017.74.0720/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1200_jco.2017.74.0720/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.0557/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.0557/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1801601/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1801601/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1513248/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1513248/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1700456/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1700456/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(19)31388-1/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(19)31388-1/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1909953/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.12618/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1708131/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1909953/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1001_jama.2019.12618/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1200_jco.2017.75.6270/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1708131/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1716995/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1716995/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1200_jco.2017.75.6270/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1708454/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(19)30772-x/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(19)30772-x/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1708454/sap.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1056_nejmoa1903869/protocol.pdf filter=lfs diff=lfs merge=lfs -text
+documents/10.1016_s0140-6736(19)31150-x/protocol.pdf filter=lfs diff=lfs merge=lfs -text
diff --git a/SYNC_README.md b/SYNC_README.md
new file mode 100644
index 0000000000000000000000000000000000000000..513bfee265d1812718014e9d929292deb7d0310b
--- /dev/null
+++ b/SYNC_README.md
@@ -0,0 +1,89 @@
+# Sync Dataset
+
+`sync_dataset.py` pulls the latest rows from the master Google Sheet, downloads the protocol/SAP PDFs for any new rows, and pushes everything to the Hugging Face dataset at [`trialdesignbench/source`](https://huggingface.co/datasets/trialdesignbench/source).
+
+## What it does
+
+1. Downloads the Google Sheet as CSV, converts to Parquet → overwrites `data/tdr.parquet`.
+2. Diffs against the previous `tdr.parquet` by the `#` column to find new rows.
+3. For each new row, downloads:
+   - `Study Protocol Link` → `documents/<#>/protocol.pdf`
+   - `Protocol+SAP / SAP Link` → `documents/<#>/sap.pdf`
+   - Skips empty links. Skips files already on disk. Writes a `*.error.txt` on failure.
+4. Uploads the updated `tdr.parquet` and any new PDFs to Hugging Face.
+
+## One-time setup
+
+```bash
+# 1. Python deps
+pip install -U "huggingface_hub[cli]" pandas pyarrow openpyxl
+# pandas + pyarrow are required for the Parquet read/write.
+# openpyxl is only needed if the source file is an uploaded .xlsx
+# (htmlview URL) rather than a native Google Sheet.
+
+# 2. Authenticate to Hugging Face (write token from https://huggingface.co/settings/tokens)
+hf auth login
+
+# 3. Make sure the Google Sheet is shared as "Anyone with the link: Viewer"
+#    Sheet URL: https://docs.google.com/spreadsheets/d/1Vb6U9Jzigtg5hLcn4R_5REW_G84cNVk6
+```
+
+## Usage
+
+Run from the `source/` directory:
+
+```bash
+cd /Users/tengjun2/Desktop/clinical/tdb/source
+
+# Preview new rows without changing anything
+python3 sync_dataset.py --dry-run
+
+# Download new PDFs and update tdr.parquet locally, but don't upload
+python3 sync_dataset.py --no-upload
+
+# Full sync: fetch sheet, download new PDFs, upload to Hugging Face
+python3 sync_dataset.py
+```
+
+## Output layout
+
+```
+source/
+├── data/
+│   └── tdr.parquet               # latest sheet content (snappy-compressed)
+├── documents/
+│   └── <row_#>/
+│       ├── protocol.pdf
+│       ├── sap.pdf
+│       ├── protocol.error.txt    # only present on download failure
+│       └── sap.error.txt
+└── sync_dataset.py
+```
+
+## Configuration
+
+Edit constants at the top of `sync_dataset.py` if any of these change:
+
+| Variable     | Default                                            | Meaning                                  |
+| ------------ | -------------------------------------------------- | ---------------------------------------- |
+| `SHEET_ID`   | `1Vb6U9Jzigtg5hLcn4R_5REW_G84cNVk6`                | Google Sheet document ID                 |
+| `SHEET_GID`  | `0`                                                | Sheet tab GID (find it in the sheet URL) |
+| `HF_REPO`    | `trialdesignbench/source`                          | Hugging Face dataset repo                |
+| `PARQUET_PATH` | `data/tdr.parquet`                               | Local Parquet path                       |
+| `DOCS_DIR`   | `documents/`                                       | PDF output root                          |
+
+## Troubleshooting
+
+**"Google Sheet is not publicly shared"** — Open the sheet → Share → General access → "Anyone with the link" → Viewer.
+
+**A PDF row writes `*.error.txt`** — The source server rejected the request (often `HTTP 403` on paywalled publisher links like NEJM). The row stays in `tdr.parquet`; only the PDF is missing. Re-running won't retry automatically — delete the `.error.txt` and the partial file, then re-run, if you want to retry.
+
+**Upload is slow** — Throughput is bounded by your upstream bandwidth. For small incremental syncs (a handful of new files) `hf upload` is used per file. For > 200 new files the script falls back to `hf upload-large-folder` which parallelizes and is resumable.
+
+**Diff missed a row whose `#` was reused** — The script keys on the `#` column. If a row's content changed without `#` changing, it won't be detected as new. Inspect manually in that case.
+
+## Notes
+
+- The script only **adds** new rows. Rows deleted from the sheet are kept in the local Parquet file and on Hugging Face — clean those up by hand if needed.
+- `tdr.parquet` is always overwritten with the latest sheet content, so edits in the sheet (typo fixes, etc.) propagate.
+- PDFs are never re-downloaded if already present. Force a re-download by deleting the file.
diff --git a/__pycache__/sync_dataset.cpython-313.pyc b/__pycache__/sync_dataset.cpython-313.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5784e93639f0d6baaeb73ec14f2dd85a42c2e074
Binary files /dev/null and b/__pycache__/sync_dataset.cpython-313.pyc differ
diff --git a/delete_old_folders.py b/delete_old_folders.py
new file mode 100644
index 0000000000000000000000000000000000000000..166a90cd2afc15d098d0f3a05a560a2b40dff567
--- /dev/null
+++ b/delete_old_folders.py
@@ -0,0 +1,42 @@
+"""Delete the old numeric-named documents/<#>/ folders from the HF dataset."""
+
+from __future__ import annotations
+
+from huggingface_hub import HfApi
+
+REPO = "trialdesignbench/source"
+
+
+def main() -> None:
+    api = HfApi()
+    files = api.list_repo_files(REPO, repo_type="dataset")
+
+    old = [
+        f for f in files
+        if f.startswith("documents/")
+        and len(f.split("/")) >= 2
+        and f.split("/")[1].isdigit()
+    ]
+    print(f"Found {len(old)} files under numeric folders. First 5:")
+    for p in old[:5]:
+        print(" ", p)
+
+    if not old:
+        return
+
+    confirm = input(f"Delete all {len(old)} files? [y/N] ").strip().lower()
+    if confirm != "y":
+        print("Aborted.")
+        return
+
+    api.delete_files(
+        delete_patterns=old,
+        repo_id=REPO,
+        repo_type="dataset",
+        commit_message="Remove old numeric-named document folders",
+    )
+    print("Deleted.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/documents/10.1001_jama.2018.20579/sap.pdf b/documents/10.1001_jama.2018.20579/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..807c5b6687c15edbd1e61ddb08fb6a2e0eee66c4
Binary files /dev/null and b/documents/10.1001_jama.2018.20579/sap.pdf differ
diff --git a/documents/10.1001_jama.2019.0071/protocol.pdf b/documents/10.1001_jama.2019.0071/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..a7f716235732ccbc0e7a6675b873db341c905791
--- /dev/null
+++ b/documents/10.1001_jama.2019.0071/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5dfe5a197820387b23f1d4570a4a1519321918afbc21f3fa5bd8a27a9cb04cc0
+size 2754354
diff --git a/documents/10.1001_jama.2019.0071/sap.pdf b/documents/10.1001_jama.2019.0071/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..594cf9a523a6774ba2d596fe8795a2e905de735d
--- /dev/null
+++ b/documents/10.1001_jama.2019.0071/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fd76358b47d339576b018881d70bada62c38e20b8a01c21e5686453ffadf4a4
+size 281032
diff --git a/documents/10.1001_jama.2019.0557/protocol.pdf b/documents/10.1001_jama.2019.0557/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..2ca77dcf9fa664aa5516db77ccd91562a66f1018
--- /dev/null
+++ b/documents/10.1001_jama.2019.0557/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08f104e10f8ca44d048262aab02bf11f63f0b59cae9d55f5fae8ff670ce79f27
+size 1209003
diff --git a/documents/10.1001_jama.2019.0557/sap.pdf b/documents/10.1001_jama.2019.0557/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..2ca77dcf9fa664aa5516db77ccd91562a66f1018
--- /dev/null
+++ b/documents/10.1001_jama.2019.0557/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08f104e10f8ca44d048262aab02bf11f63f0b59cae9d55f5fae8ff670ce79f27
+size 1209003
diff --git a/documents/10.1001_jama.2019.12618/protocol.pdf b/documents/10.1001_jama.2019.12618/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3cb9f9975b656a00144711682f1a5a70d8595373
--- /dev/null
+++ b/documents/10.1001_jama.2019.12618/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e0d9fff5fb3b8681a408213bc82810107ea8689d4e2a7f0c5ddd060fc98193
+size 9456599
diff --git a/documents/10.1001_jama.2019.12618/sap.pdf b/documents/10.1001_jama.2019.12618/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3cb9f9975b656a00144711682f1a5a70d8595373
--- /dev/null
+++ b/documents/10.1001_jama.2019.12618/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e0d9fff5fb3b8681a408213bc82810107ea8689d4e2a7f0c5ddd060fc98193
+size 9456599
diff --git a/documents/10.1001_jama.2019.1660/protocol.pdf b/documents/10.1001_jama.2019.1660/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..7836c4ea1fbcda1a626cd322df1273e1ff5e2d81
--- /dev/null
+++ b/documents/10.1001_jama.2019.1660/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96abe28b22bfb56939c150fc143d024d306d81fca255f54550ce8cb194e2f5cc
+size 1215290
diff --git a/documents/10.1001_jama.2019.1660/sap.pdf b/documents/10.1001_jama.2019.1660/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0fe788edae59c11284363705c7e7177cecf35452
--- /dev/null
+++ b/documents/10.1001_jama.2019.1660/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f9ba880010ccd6a2581715b5824c4759f437337d36372bf800b4c307245b549
+size 408047
diff --git a/documents/10.1016_s0140-6736(17)33297-x/protocol.pdf b/documents/10.1016_s0140-6736(17)33297-x/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c0654b4513159c995bc2d46eab0d989e92b07191
--- /dev/null
+++ b/documents/10.1016_s0140-6736(17)33297-x/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f64a224e34e8b1ccab7161b025b8944ff05f0e06e15c4ba70d14b1713daccb3
+size 2491616
diff --git a/documents/10.1016_s0140-6736(17)33297-x/sap.pdf b/documents/10.1016_s0140-6736(17)33297-x/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c0654b4513159c995bc2d46eab0d989e92b07191
--- /dev/null
+++ b/documents/10.1016_s0140-6736(17)33297-x/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f64a224e34e8b1ccab7161b025b8944ff05f0e06e15c4ba70d14b1713daccb3
+size 2491616
diff --git a/documents/10.1016_s0140-6736(18)32321-3/protocol.pdf b/documents/10.1016_s0140-6736(18)32321-3/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..93ea1b219bafcfd83e98edcb1cb48ee65101e600
--- /dev/null
+++ b/documents/10.1016_s0140-6736(18)32321-3/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f0c1405283c72390c92ca645b3f62502b18ba7b7aebd59ea6bf3f25e5e51395
+size 1079157
diff --git a/documents/10.1016_s0140-6736(18)32321-3/sap.pdf b/documents/10.1016_s0140-6736(18)32321-3/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..93ea1b219bafcfd83e98edcb1cb48ee65101e600
--- /dev/null
+++ b/documents/10.1016_s0140-6736(18)32321-3/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f0c1405283c72390c92ca645b3f62502b18ba7b7aebd59ea6bf3f25e5e51395
+size 1079157
diff --git a/documents/10.1016_s0140-6736(19)30772-x/protocol.pdf b/documents/10.1016_s0140-6736(19)30772-x/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..ee2f554ef0719aecfc07acac219223b58d74a9b1
--- /dev/null
+++ b/documents/10.1016_s0140-6736(19)30772-x/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0098f38b6a5831341ce5f5ef1a8a2b27bebed3c5fee846ca099bb6ce690457ce
+size 2004321
diff --git a/documents/10.1016_s0140-6736(19)30772-x/sap.pdf b/documents/10.1016_s0140-6736(19)30772-x/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3b304082a7e8adc788eab9004fa1babe023a393e
--- /dev/null
+++ b/documents/10.1016_s0140-6736(19)30772-x/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f9d0f965bc28eafc1229d293318c6ab1b57112f55411ad092bbeb147b082a41
+size 795974
diff --git a/documents/10.1016_s0140-6736(19)31150-x/protocol.pdf b/documents/10.1016_s0140-6736(19)31150-x/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..e61663f5e4b1f95568026cd4dccfcf3d8ab2d8bf
--- /dev/null
+++ b/documents/10.1016_s0140-6736(19)31150-x/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c379993f99707de410d645765a4282e966085eef5379847d029c6c592884c40
+size 918504
diff --git a/documents/10.1016_s0140-6736(19)31388-1/protocol.pdf b/documents/10.1016_s0140-6736(19)31388-1/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..94c381f3eeb6347df8c04e09c7162d808dd76d38
--- /dev/null
+++ b/documents/10.1016_s0140-6736(19)31388-1/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:247060db1ac1a71eb649698297fde5ecf50dd927249822c001dafda788b95545
+size 1446328
diff --git a/documents/10.1016_s0140-6736(19)31388-1/sap.pdf b/documents/10.1016_s0140-6736(19)31388-1/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..d9fd5850a6957dd1bdfaf0a42d64ad8560cba452
--- /dev/null
+++ b/documents/10.1016_s0140-6736(19)31388-1/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0444958cf55da7f5d82d21ffc0809f728a7b77707d32c9a1257ac75e8c2f676
+size 1054189
diff --git a/documents/10.1016_s0140-6736(19)31793-3/sap.pdf b/documents/10.1016_s0140-6736(19)31793-3/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0d1f888ff78fa00072677a06494f2bfad90b17a7
Binary files /dev/null and b/documents/10.1016_s0140-6736(19)31793-3/sap.pdf differ
diff --git a/documents/10.1056_nejmoa1513248/protocol.pdf b/documents/10.1056_nejmoa1513248/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..d3794a14b36fa2e48829cfbe00ba10f99d8cda5f
--- /dev/null
+++ b/documents/10.1056_nejmoa1513248/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8d22817a7019ded0b8998c9d8578e444ec91af75355592108b5a6b72f10cbc3
+size 872936
diff --git a/documents/10.1056_nejmoa1513248/sap.pdf b/documents/10.1056_nejmoa1513248/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..053cebd56aef904076ae4897551f5b36ac278c31
--- /dev/null
+++ b/documents/10.1056_nejmoa1513248/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a646ff761268b9cc00b6762529f91d4ea93c7b39af5e4f2b152b46f52a2d47b
+size 672972
diff --git a/documents/10.1056_nejmoa1700456/protocol.pdf b/documents/10.1056_nejmoa1700456/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4843753494b0b9bcf712f36594f672a98c9e2601
--- /dev/null
+++ b/documents/10.1056_nejmoa1700456/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37a4853dffc6a551580d6c14b40179c8ee8bff8fbbcd2f2280ebd9c9a253941b
+size 13460546
diff --git a/documents/10.1056_nejmoa1700456/sap.pdf b/documents/10.1056_nejmoa1700456/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4843753494b0b9bcf712f36594f672a98c9e2601
--- /dev/null
+++ b/documents/10.1056_nejmoa1700456/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37a4853dffc6a551580d6c14b40179c8ee8bff8fbbcd2f2280ebd9c9a253941b
+size 13460546
diff --git a/documents/10.1056_nejmoa1708131/protocol.pdf b/documents/10.1056_nejmoa1708131/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..17effd444a8145040660474e1c1e315f91d9513a
--- /dev/null
+++ b/documents/10.1056_nejmoa1708131/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:764d9df5bee793e75b0251679b8585674b4913dd21ac4042210a14039537e194
+size 4316285
diff --git a/documents/10.1056_nejmoa1708131/sap.pdf b/documents/10.1056_nejmoa1708131/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c200fb87aa988ce6845462755d6622e9627fe439
--- /dev/null
+++ b/documents/10.1056_nejmoa1708131/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aecd4b62ebda8a34ea4d8aebf0b3f689336f4205d02351c1a3fa892d5c8571d
+size 260398
diff --git a/documents/10.1056_nejmoa1708454/protocol.pdf b/documents/10.1056_nejmoa1708454/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c92659eae57485117db762b7b6308258b5d39cb6
--- /dev/null
+++ b/documents/10.1056_nejmoa1708454/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94d6d58fa1594954193d172da6b1ffc948b83decef55edcd9cf7e99a0ae91181
+size 1202771
diff --git a/documents/10.1056_nejmoa1708454/sap.pdf b/documents/10.1056_nejmoa1708454/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3f6920e7a1e15a36944eccb5441acff3c0214249
--- /dev/null
+++ b/documents/10.1056_nejmoa1708454/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02d15b9ea9d1a4913484e553801be2ca633ff02031f0be63d946a6d54c0915a1
+size 4557136
diff --git a/documents/10.1056_nejmoa1716995/protocol.pdf b/documents/10.1056_nejmoa1716995/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..33df75623e7079f545973c47bd932d8362685b62
--- /dev/null
+++ b/documents/10.1056_nejmoa1716995/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:486a7c300a9a3c8a8a4d027f1d8bff151722e715d762c406230b32dcbd7c5aa5
+size 1195199
diff --git a/documents/10.1056_nejmoa1716995/sap.pdf b/documents/10.1056_nejmoa1716995/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..cf67fa2e042be67b71c18be47f0e1c255ea2fa9b
--- /dev/null
+++ b/documents/10.1056_nejmoa1716995/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f94679ff396d4becff0380b0c9b6977cf633d4a0167244355d4d29009afbde84
+size 851897
diff --git a/documents/10.1056_nejmoa1801601/protocol.pdf b/documents/10.1056_nejmoa1801601/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..859963cb3fd1fb4d6619aa67a08ad11ebe8dd02b
--- /dev/null
+++ b/documents/10.1056_nejmoa1801601/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d270558bb30306f8d965f5826d3debf14d087261934ecc74516e39b973332880
+size 1080077
diff --git a/documents/10.1056_nejmoa1801601/sap.pdf b/documents/10.1056_nejmoa1801601/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3cd1ff0ae923ddc3c0d5ed6cc2db4b51d04368a4
--- /dev/null
+++ b/documents/10.1056_nejmoa1801601/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b3276071d3ebd13e65f2d04b0d6bb3e4e9539963a6437d86bcc3aac6353a57f
+size 166155
diff --git a/documents/10.1056_nejmoa1805104/protocol.pdf b/documents/10.1056_nejmoa1805104/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..58fb9e73ccfe301f5df3cc98e5443ed6cd66f3e7
--- /dev/null
+++ b/documents/10.1056_nejmoa1805104/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9ba3470bc0bb95641699f1fe08359ec229d72d25999738bee206c49f1be4005
+size 6501069
diff --git a/documents/10.1056_nejmoa1805104/sap.pdf b/documents/10.1056_nejmoa1805104/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..58fb9e73ccfe301f5df3cc98e5443ed6cd66f3e7
--- /dev/null
+++ b/documents/10.1056_nejmoa1805104/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9ba3470bc0bb95641699f1fe08359ec229d72d25999738bee206c49f1be4005
+size 6501069
diff --git a/documents/10.1056_nejmoa1816714/protocol.pdf b/documents/10.1056_nejmoa1816714/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5f455dd15f42124593f96e363b210b491fe38e07
--- /dev/null
+++ b/documents/10.1056_nejmoa1816714/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14496801a4397f97ced0c949623550da93a5d8cb9a17982b9a8791cecce64915
+size 2578319
diff --git a/documents/10.1056_nejmoa1816714/sap.pdf b/documents/10.1056_nejmoa1816714/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..5f455dd15f42124593f96e363b210b491fe38e07
--- /dev/null
+++ b/documents/10.1056_nejmoa1816714/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14496801a4397f97ced0c949623550da93a5d8cb9a17982b9a8791cecce64915
+size 2578319
diff --git a/documents/10.1056_nejmoa1903869/protocol.pdf b/documents/10.1056_nejmoa1903869/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..04e49cf33dc1045dba4a322ddf0002a37522d3a3
--- /dev/null
+++ b/documents/10.1056_nejmoa1903869/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a765ac16832c3638653c44c89cf033d8a5a55b80855af0818378e9bc9e2212fd
+size 2184520
diff --git a/documents/10.1056_nejmoa1908655/protocol.pdf b/documents/10.1056_nejmoa1908655/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..2ceae60b2853fb208c2ae5e35577d8c351c7a47a
--- /dev/null
+++ b/documents/10.1056_nejmoa1908655/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c645095faa926776d2f9f9f564dd348d44e348d4c3105d36a20276d33ea904b7
+size 1231466
diff --git a/documents/10.1056_nejmoa1908655/sap.pdf b/documents/10.1056_nejmoa1908655/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..9562b6b7ed6892f4d8d8f8a8855d5d1790c440aa
--- /dev/null
+++ b/documents/10.1056_nejmoa1908655/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5590f7a09d944d30e6b5f31b78fa67e8982fda1c92d9eaedbca87c025b2bdb0f
+size 601155
diff --git a/documents/10.1056_nejmoa1909953/protocol.pdf b/documents/10.1056_nejmoa1909953/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..a03e2636b505c57828125e5e749cfd0ea6b1ace6
--- /dev/null
+++ b/documents/10.1056_nejmoa1909953/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d4fe3f2632832fb7695159422d10aff7a4df6d841777fe34bd644abd4e787ae
+size 2297612
diff --git a/documents/10.1056_nejmoa1909953/sap.pdf b/documents/10.1056_nejmoa1909953/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..3c1dde1a0d4870056fe0a59d503576e4cd38da3f
--- /dev/null
+++ b/documents/10.1056_nejmoa1909953/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd81a8db29ff12599057679d1c59345d81c5455387385b65a310a7a006203315
+size 562980
diff --git a/documents/10.1200_jco.2017.74.0720/protocol.pdf b/documents/10.1200_jco.2017.74.0720/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..0ccce87fa8154f6b1718583f5882edf36dfefbb5
--- /dev/null
+++ b/documents/10.1200_jco.2017.74.0720/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:511f2be5a0dde243fa66512139c1029e0a6032db84c4fbc5cecbed7a99f887a6
+size 478930
diff --git a/documents/10.1200_jco.2017.74.0720/sap.pdf b/documents/10.1200_jco.2017.74.0720/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..31429bb13154401bb1bd3c762d0a623fc60963a1
--- /dev/null
+++ b/documents/10.1200_jco.2017.74.0720/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:003097bc03e2266b9476c5c48d1064e0b24809fa7c8019855358a13a26c2795a
+size 359491
diff --git a/documents/10.1200_jco.2017.75.6270/protocol.pdf b/documents/10.1200_jco.2017.75.6270/protocol.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4b1d2ccfbfb933fba79b787287b9d0c0351cbf71
--- /dev/null
+++ b/documents/10.1200_jco.2017.75.6270/protocol.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a312bf6befdd59bcaf8edbce725b215636d075fb52437c59286a2c1e4a3f2ffa
+size 3378050
diff --git a/documents/10.1200_jco.2017.75.6270/sap.pdf b/documents/10.1200_jco.2017.75.6270/sap.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4b1d2ccfbfb933fba79b787287b9d0c0351cbf71
--- /dev/null
+++ b/documents/10.1200_jco.2017.75.6270/sap.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a312bf6befdd59bcaf8edbce725b215636d075fb52437c59286a2c1e4a3f2ffa
+size 3378050
diff --git a/download_docs.py b/download_docs.py
index b6f6d33f9ec2f19018185ba55ef8b8824e4bbd8e..6355debcad20f086fc3f96baaa685a33bdc851ec 100644
--- a/download_docs.py
+++ b/download_docs.py
@@ -12,6 +12,7 @@ import time
 import urllib.error
 import urllib.request
 from pathlib import Path
+from urllib.parse import urlparse
 
 CSV_PATH = Path(__file__).parent / "file.csv"
 OUT_ROOT = Path(__file__).parent / "downloads"
@@ -25,6 +26,12 @@ UA = (
 )
 
 
+def paper_link_slug(link: str) -> str:
+    path = urlparse((link or "").strip()).path.strip("/")
+    parts = [p for p in path.split("/") if p]
+    return "_".join(parts[-2:]) if len(parts) >= 2 else ""
+
+
 def download(url: str, dest: Path) -> tuple[bool, str]:
     if dest.exists() and dest.stat().st_size > 0:
         return True, "exists"
@@ -64,7 +71,12 @@ def main(start_row: int = 3) -> None:
                 print(f"[{num}] skip: no links")
                 continue
 
-            row_dir = OUT_ROOT / str(num)
+            slug = paper_link_slug(row.get("Paper Link") or "")
+            if not slug:
+                print(f"[{num}] skip: no usable Paper Link")
+                continue
+
+            row_dir = OUT_ROOT / slug
             row_dir.mkdir(exist_ok=True)
 
             if protocol:
diff --git a/replace_doc.py b/replace_doc.py
new file mode 100644
index 0000000000000000000000000000000000000000..65f5218157056e3cc21f849d426444c76d99d67f
--- /dev/null
+++ b/replace_doc.py
@@ -0,0 +1,15 @@
+from huggingface_hub import HfApi
+
+api = HfApi()
+# api.delete_folder(
+#     "documents",
+#     repo_id="trialdesignbench/source",
+#     repo_type="dataset",
+# )
+api.upload_large_folder(
+    folder_path="/users/teng77/source",     # parent — preserves "documents/" prefix
+    repo_id="trialdesignbench/source",
+    repo_type="dataset",
+    allow_patterns=["documents/**"],
+    ignore_patterns=[".cache/**", "**/.cache/**"],
+)
diff --git a/sync_dataset.py b/sync_dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..3cb2a2ef945174c8d210dc881f3451c95ab6afaa
--- /dev/null
+++ b/sync_dataset.py
@@ -0,0 +1,323 @@
+"""Sync TrialDesignBench dataset from Google Sheets to local + Hugging Face.
+
+Steps:
+1. Download the latest sheet as CSV from Google Sheets.
+2. Diff against the existing tdr.parquet by the "#" column to find new rows.
+3. Download protocol and SAP PDFs for new rows (skip if no link).
+4. Overwrite tdr.parquet and upload the changed files to Hugging Face.
+
+Usage:
+    python sync_dataset.py              # full sync
+    python sync_dataset.py --no-upload  # local only
+    python sync_dataset.py --dry-run    # show what would happen
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+from urllib.parse import urlparse
+
+SHEET_ID = "1Vb6U9Jzigtg5hLcn4R_5REW_G84cNVk6"
+SHEET_GID = "0"
+SHEET_CSV_URL = (
+    f"https://docs.google.com/spreadsheets/d/{SHEET_ID}/export"
+    f"?format=csv&gid={SHEET_GID}"
+)
+# Fallback for uploaded .xlsx files (htmlview URLs) — downloads raw bytes
+# and converts to CSV via openpyxl.
+SHEET_XLSX_URL = f"https://docs.google.com/uc?export=download&id={SHEET_ID}"
+SHEET_TAB_INDEX = 0  # 0 = first sheet; change if data is on another tab
+
+ROOT = Path(__file__).parent
+PARQUET_PATH = ROOT / "data" / "tdr.parquet"
+DOCS_DIR = ROOT / "documents"
+HF_REPO = "trialdesignbench/source"
+
+UA = (
+    "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
+    "AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
+)
+TIMEOUT = 60
+RETRIES = 2
+
+
+def fetch_sheet_csv() -> str:
+    """Try native-sheet CSV export first; fall back to xlsx download + convert."""
+    try:
+        req = urllib.request.Request(SHEET_CSV_URL, headers={"User-Agent": UA})
+        with urllib.request.urlopen(req, timeout=TIMEOUT) as resp:
+            final_url = resp.geturl()
+            if "accounts.google.com" in final_url or "ServiceLogin" in final_url:
+                raise RuntimeError(
+                    "Google Sheet is not publicly shared — "
+                    "enable 'Anyone with link: Viewer'."
+                )
+            return resp.read().decode("utf-8")
+    except urllib.error.HTTPError as e:
+        if e.code != 400:
+            raise
+        print("CSV export returned 400 — falling back to .xlsx download.")
+        return _fetch_xlsx_as_csv()
+
+
+def _fetch_xlsx_as_csv() -> str:
+    try:
+        from openpyxl import load_workbook
+    except ImportError as e:
+        msg = (
+            "openpyxl is required to read uploaded .xlsx Drive files. "
+            "Install with: pip install openpyxl"
+        )
+        raise RuntimeError(msg) from e
+
+    import io
+
+    req = urllib.request.Request(SHEET_XLSX_URL, headers={"User-Agent": UA})
+    with urllib.request.urlopen(req, timeout=TIMEOUT) as resp:
+        final_url = resp.geturl()
+        if "accounts.google.com" in final_url or "ServiceLogin" in final_url:
+            raise RuntimeError(
+                "File is not publicly shared — enable 'Anyone with link: Viewer'."
+            )
+        data = resp.read()
+
+    wb = load_workbook(io.BytesIO(data), read_only=True, data_only=True)
+    ws = wb.worksheets[SHEET_TAB_INDEX]
+
+    def _fmt(v: object) -> str:
+        if v is None:
+            return ""
+        # openpyxl returns whole-number cells as floats (1.0, 2.0, ...);
+        # collapse back to int so "#" / "Year" / "PMID" stay digit-like.
+        if isinstance(v, float) and v.is_integer():
+            return str(int(v))
+        return str(v)
+
+    buf = io.StringIO()
+    writer = csv.writer(buf)
+    for row in ws.iter_rows(values_only=True):
+        writer.writerow([_fmt(v) for v in row])
+    return buf.getvalue()
+
+
+def read_existing_ids() -> set[str]:
+    if not PARQUET_PATH.exists():
+        return set()
+    import pandas as pd
+
+    df = pd.read_parquet(PARQUET_PATH, columns=["#"])
+    out: set[str] = set()
+    for v in df["#"].dropna():
+        # Normalize whole-number floats ("1.0") back to "1" so the diff
+        # matches sheet rows regardless of how the column was stored.
+        if isinstance(v, float) and v.is_integer():
+            out.add(str(int(v)))
+        else:
+            s = str(v).strip()
+            if s:
+                out.add(s)
+    return out
+
+
+def write_parquet(csv_text: str) -> None:
+    import io
+
+    import pandas as pd
+
+    df = pd.read_csv(io.StringIO(csv_text))
+    PARQUET_PATH.parent.mkdir(parents=True, exist_ok=True)
+    df.to_parquet(PARQUET_PATH, index=False, compression="snappy")
+
+
+def parse_rows(csv_text: str) -> list[dict[str, str]]:
+    reader = csv.DictReader(csv_text.splitlines())
+    return [row for row in reader if (row.get("#") or "").strip().isdigit()]
+
+
+def download_pdf(url: str, dest: Path) -> tuple[bool, str]:
+    if dest.exists() and dest.stat().st_size > 0:
+        return True, "exists"
+    req = urllib.request.Request(url, headers={"User-Agent": UA})
+    last_err = ""
+    for attempt in range(RETRIES + 1):
+        try:
+            with urllib.request.urlopen(req, timeout=TIMEOUT) as resp:
+                data = resp.read()
+            dest.write_bytes(data)
+            return True, f"ok ({len(data)} bytes)"
+        except urllib.error.HTTPError as e:
+            last_err = f"HTTP {e.code}"
+        except (urllib.error.URLError, TimeoutError) as e:
+            last_err = f"network: {e}"
+        except Exception as e:  # noqa: BLE001
+            last_err = f"error: {e}"
+        time.sleep(1 + attempt)
+    return False, last_err
+
+
+def paper_link_slug(link: str) -> str:
+    """Last two path segments of a Paper Link joined with '_'.
+
+    Example: https://doi.org/10.1056/nejmoa2511478 -> 10.1056_nejmoa2511478
+    """
+    path = urlparse((link or "").strip()).path.strip("/")
+    parts = [p for p in path.split("/") if p]
+    return "_".join(parts[-2:]) if len(parts) >= 2 else ""
+
+
+def _row_missing_pdfs(row: dict[str, str]) -> bool:
+    """A row needs work iff it has a usable slug + at least one link whose
+    target PDF is not already on disk."""
+    slug = paper_link_slug(row.get("Paper Link") or "")
+    if not slug:
+        return False
+    protocol = (row.get("Study Protocol Link") or "").strip()
+    sap = (row.get("Protocol+SAP / SAP Link") or "").strip()
+    if not protocol and not sap:
+        return False
+    row_dir = DOCS_DIR / slug
+    if protocol and not (row_dir / "protocol.pdf").exists():
+        return True
+    if sap and not (row_dir / "sap.pdf").exists():
+        return True
+    return False
+
+
+def download_for_row(row: dict[str, str]) -> list[Path]:
+    num = (row.get("#") or "").strip()
+    slug = paper_link_slug(row.get("Paper Link") or "")
+    protocol = (row.get("Study Protocol Link") or "").strip()
+    sap = (row.get("Protocol+SAP / SAP Link") or "").strip()
+    if not protocol and not sap:
+        print(f"[{num}] skip: no links")
+        return []
+    if not slug:
+        print(f"[{num}] skip: no usable Paper Link for folder name")
+        return []
+
+    row_dir = DOCS_DIR / slug
+    row_dir.mkdir(parents=True, exist_ok=True)
+    new_files: list[Path] = []
+
+    if protocol:
+        dest = row_dir / "protocol.pdf"
+        existed = dest.exists()
+        ok, msg = download_pdf(protocol, dest)
+        print(f"[{num}/{slug}] protocol: {msg}")
+        if ok and not existed:
+            new_files.append(dest)
+        elif not ok:
+            (row_dir / "protocol.error.txt").write_text(
+                f"{protocol}\n{msg}\n", encoding="utf-8"
+            )
+
+    if sap:
+        dest = row_dir / "sap.pdf"
+        existed = dest.exists()
+        ok, msg = download_pdf(sap, dest)
+        print(f"[{num}/{slug}] sap: {msg}")
+        if ok and not existed:
+            new_files.append(dest)
+        elif not ok:
+            (row_dir / "sap.error.txt").write_text(
+                f"{sap}\n{msg}\n", encoding="utf-8"
+            )
+
+    return new_files
+
+
+def _hf_cli() -> str:
+    # Resolve the `hf` CLI even when PATH doesn't include the active venv/conda env
+    # (e.g. when this script is launched from a non-interactive shell).
+    import shutil
+
+    found = shutil.which("hf")
+    if found:
+        return found
+    candidate = Path(sys.executable).parent / "hf"
+    if candidate.exists():
+        return str(candidate)
+    raise RuntimeError(
+        "`hf` CLI not found. Install with: pip install -U 'huggingface_hub[cli]'"
+    )
+
+
+def hf_upload(paths: list[Path]) -> None:
+    if not paths:
+        print("No files to upload.")
+        return
+    rels = [str(p.relative_to(ROOT)) for p in paths]
+    print(f"Uploading {len(rels)} files to {HF_REPO} ...")
+    hf = _hf_cli()
+    # Use hf upload for small incremental batches; switch to upload-large-folder
+    # if the new-row set is large.
+    if len(rels) > 200:
+        cmd = [
+            hf, "upload-large-folder", HF_REPO, str(ROOT),
+            "--repo-type=dataset", "--num-workers=4",
+        ]
+        subprocess.run(cmd, check=True)
+        return
+    for rel in rels:
+        cmd = [hf, "upload", HF_REPO, rel, rel, "--repo-type=dataset"]
+        subprocess.run(cmd, check=True)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--no-upload", action="store_true", help="Skip HF upload.")
+    parser.add_argument(
+        "--dry-run", action="store_true", help="Show diff only, no downloads or upload."
+    )
+    args = parser.parse_args()
+
+    print(f"Fetching sheet ({SHEET_ID}, gid={SHEET_GID}) ...")
+    csv_text = fetch_sheet_csv()
+    new_rows_all = parse_rows(csv_text)
+    print(f"Sheet has {len(new_rows_all)} data rows.")
+
+    existing_ids = read_existing_ids()
+    print(f"Local parquet has {len(existing_ids)} rows.")
+
+    new_in_sheet = [r for r in new_rows_all if (r.get("#") or "").strip() not in existing_ids]
+    rows_needing_pdfs = [r for r in new_rows_all if _row_missing_pdfs(r)]
+    print(f"New rows in sheet (vs parquet #): {len(new_in_sheet)}")
+    print(f"Rows missing PDFs on disk: {len(rows_needing_pdfs)}")
+    for r in rows_needing_pdfs[:20]:
+        print(f"  - #{r.get('#')}: {(r.get('Paper Title') or '')[:80]}")
+    if len(rows_needing_pdfs) > 20:
+        print(f"  ... and {len(rows_needing_pdfs) - 20} more")
+
+    if args.dry_run:
+        return
+
+    write_parquet(csv_text)
+    print(f"Wrote {PARQUET_PATH}")
+
+    DOCS_DIR.mkdir(exist_ok=True)
+    new_pdfs: list[Path] = []
+    for row in rows_needing_pdfs:
+        new_pdfs.extend(download_for_row(row))
+
+    print(f"Downloaded {len(new_pdfs)} new PDFs.")
+
+    if args.no_upload:
+        return
+
+    hf_upload([PARQUET_PATH, *new_pdfs])
+    print("Done.")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except KeyboardInterrupt:
+        print("\nAborted.")
+        sys.exit(130)