add video shards

README.md

@@ -68,6 +68,9 @@ dataset_info:
   dataset_size: 413638
 ---
 
+> [!IMPORTANT]
+> This is a fork of the original [`yale-nlp/TOMATO`](https://huggingface.co/datasets/yale-nlp/TOMATO) HF dataset. We have uploaded the `videos.zip` file directly here so that eval with `lmms-eval` works right out of the box without running a separate process.
+
 # 🍅 TOMATO
 
 [**📄 Paper**](https://arxiv.org/abs/2410.23266) | [**💻 Code**](https://github.com/yale-nlp/TOMATO) | [**🎬 Videos**](https://drive.google.com/file/d/1-dNt9bZcp6C3RXuGoAO3EBgWkAHg8NWR/view?usp=drive_link) 

tar_videos.py

@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.10"
+# ///
+"""
+Tar the videos in the TOMATO dataset into shards of 5 GiB each
+"""
+
+import subprocess
+from pathlib import Path
+import random
+
+# ---- Config ----
+root =  Path(__file__).parent
+seed = 42
+TARGET_MAX_GB    = 2.25                              # hard cap: 2.25 GiB
+TARGET_MAX_BYTES = TARGET_MAX_GB * 1024**3           # hard cap
+SAFETY_MARGIN    = 256 * 1024**2                     # keep ~256 MiB headroom
+PACK_LIMIT       = TARGET_MAX_BYTES - SAFETY_MARGIN  # packing budget per shard
+SHARD_PREFIX     = "video_shard"                     # name prefix
+ZSTD_ARGS        = "zstd -T0 -3"                     # multithreaded, modest level
+DRY_RUN          = False                             # set True to preview
+EXTENSIONS       = [".mp4"]                          # video extensions
+
+# ---- 1) Collect & shuffle globally ----
+# Adjust the filter to match only video extensions if needed (mp4, mkv, etc.)
+print(f"Collecting video files from {root}...")
+video_files = sorted(root.rglob("videos/*/*.mp4"))
+print(f"Found {len(video_files)} files")
+
+# filter to only video files
+video_files = [f for f in video_files if f.suffix.lower() in EXTENSIONS]
+print(f"Filtered to {len(video_files)} video files")
+
+random.seed(seed)
+random.shuffle(video_files)
+
+def human(n):
+    for u in ["B","KiB","MiB","GiB","TiB"]:
+        if n < 1024 or u == "TiB": return f"{n:.1f} {u}"
+        n /= 1024
+
+# ---- 2) Greedy size-based packing ----
+shards = []
+cur_list, cur_bytes = [], 0
+
+for f in video_files:
+    sz = f.stat().st_size
+    # If a single file exceeds PACK_LIMIT, put it alone to respect 5 GiB cap
+    if sz > PACK_LIMIT:
+        if cur_list:
+            shards.append(cur_list)
+            cur_list, cur_bytes = [], 0
+        shards.append([f])
+        continue
+
+    if cur_bytes + sz > PACK_LIMIT and cur_list:
+        shards.append(cur_list)
+        cur_list, cur_bytes = [f], sz
+    else:
+        cur_list.append(f)
+        cur_bytes += sz
+
+if cur_list:
+    shards.append(cur_list)
+
+print(f"Planned {len(shards)} shards")
+for i, s in enumerate(shards):
+    total = sum(p.stat().st_size for p in s)
+    print(f"  {i:03d}: {len(s)} files, ~{human(total)} uncompressed")
+
+# ---- 3) Materialize shards with tar -T and zstd ----
+for i, shard in enumerate(shards):
+    listfile = root / f".tarlist_{i:03d}.txt"
+    listfile.write_text("\n".join(str(p.relative_to(root)) for p in shard))
+
+    tar_name = f"{SHARD_PREFIX}_{i:03d}.tar.zst"
+    tar_path = root / tar_name
+
+    cmd = [
+        "tar",
+        "--use-compress-program", ZSTD_ARGS,
+        "-cf", str(tar_path),
+        "-C", str(root),
+        "--checkpoint=1000",
+        "--checkpoint-action=dot",
+        "-T", str(listfile),
+    ]
+    print(f"\nCreating {tar_name} with {len(shard)} files...")
+    print("Command:", " ".join(cmd))
+
+    if not DRY_RUN:
+        # shell=True so --use-compress-program arg is parsed correctly
+        subprocess.run(cmd, check=True)
+        listfile.unlink(missing_ok=True)
+        # Optional: hard assert size cap
+        out_sz = tar_path.stat().st_size
+        if out_sz > TARGET_MAX_BYTES:
+            raise RuntimeError(
+                f"{tar_name} is {human(out_sz)} which exceeds the 5 GiB cap"
+            )
+    else:
+        print("DRY RUN — not executing tar")

video_shard_000.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad8306ee0fa091c93a0bd5df68a08a905a34cfb84e275fac188084b77c3b56e9
+size 2116126950

video_shard_001.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a2830777a10216d85428d8b455a2535d7b81dc327b840cf2920ecebf100c09a
+size 2091538640

video_shard_002.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8fd07c2a11a2c9a5dc60edb5cbf06fe6166e54be6c9d0f8383c083104471114
+size 2132517052

video_shard_003.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de35cd0eff41db8323712500edf67dd0e6224f7caf99f832de8ccf06a5616171
+size 2106316428

video_shard_004.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae26854d69924a24a0ed9eaef291aab1101c579fc098a535008c3d528bb50258
+size 2128406059

video_shard_005.tar.zst

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99c76cc03e44e54b83a4fbc7c66c294ad06c0378618e46eb51eba55a01483002
+size 1452639983