Add 10s camera segment dataset

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+sekai_game_train_10s_camera_segments/manifest.jsonl filter=lfs diff=lfs merge=lfs -text

sekai_game_train_10s_camera_segments/README_zh.md

@@ -0,0 +1,35 @@
+# sekai_game_train_10s_camera_segments
+
+由 `/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/haozhu/data/sekai_game_train_961frames_16fps_ovl640` 整理得到。该版本不包含原始视频；每条样本只包含一张输入图、一段 prompt、以及对应 10 秒的逐帧相机参数。
+
+## 采样规则
+
+- 原始视频：1618 条，每条 961 帧，16 fps。
+- 每条视频按非重叠 10 秒片段采样：每段 160 帧。
+- 每条原视频取 6 段：`[0,160)`, `[160,320)`, `[320,480)`, `[480,640)`, `[640,800)`, `[800,960)`。
+- 每段保存该段第 1 帧作为 `input_image`，保存原视频 prompt，保存该段 160 帧的 `poses` 和 `intrinsics`。
+- 原视频多出的最后第 961 帧不单独成段。
+
+## 数据格式
+
+数据以 WebDataset 风格 tar shards 存在 `shards/` 中。每个样本包含：
+
+- `samples/<id>.jpg`：10 秒片段首帧图像。
+- `samples/<id>.txt`：prompt。
+- `samples/<id>.npz`：包含 `poses (160, 4, 4)` 和 `intrinsics (160, 4)`。
+- `samples/<id>.json`：元数据。
+
+根目录文件：
+
+- `manifest.jsonl`：每条样本一行，记录 shard、tar member、source key、segment index、frame range 等。
+- `stats.json`：整理统计。
+
+## 统计
+
+- source videos: 1618
+- segments per video: 6
+- total samples: 9708
+- segment frames: 160
+- segment seconds: 10
+- fps: 16.0
+- shards: 17

sekai_game_train_10s_camera_segments/manifest.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5de1cf9f8d4665447464c0b667e5fbdd6d9fe0b030bb738476f01922cd80528
+size 14534914

sekai_game_train_10s_camera_segments/prepare_sekai_10s_camera_dataset.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+"""Prepare 10-second camera-control samples from sekai_game_train.
+
+Each output sample is stored in WebDataset-style tar shards with:
+  - <sample_id>.jpg: first frame of the 10s segment
+  - <sample_id>.txt: prompt
+  - <sample_id>.npz: poses (160, 4, 4) and intrinsics (160, 4)
+  - <sample_id>.json: metadata
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import io
+import json
+import multiprocessing as mp
+import os
+import shutil
+import subprocess
+import tarfile
+import tempfile
+import time
+import zipfile
+from pathlib import Path
+
+import imageio_ffmpeg
+import numpy as np
+
+
+SRC_DIR = Path(
+    "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/haozhu/data/"
+    "sekai_game_train_961frames_16fps_ovl640"
+)
+ZIP_PATH = SRC_DIR / "sekai_game_train_00000000.zip"
+CAMERA_PATH = SRC_DIR / "sekai_game_train_00000000_camera.npz"
+
+FPS = 16.0
+SEGMENT_SECONDS = 10
+SEGMENT_FRAMES = int(FPS * SEGMENT_SECONDS)
+SEGMENTS_PER_VIDEO = 6
+FRAME_STARTS = [i * SEGMENT_FRAMES for i in range(SEGMENTS_PER_VIDEO)]
+
+
+def add_bytes_to_tar(tar: tarfile.TarFile, name: str, data: bytes) -> None:
+    info = tarfile.TarInfo(name)
+    info.size = len(data)
+    info.mtime = 0
+    tar.addfile(info, io.BytesIO(data))
+
+
+def make_sample_id(key: str, segment_index: int, frame_start: int) -> str:
+    frame_end = frame_start + SEGMENT_FRAMES
+    return f"{key}_seg{segment_index:02d}_f{frame_start:04d}_{frame_end:04d}"
+
+
+def extract_segment_frames(
+    ffmpeg: str,
+    mp4_path: Path,
+    frame_dir: Path,
+    jpeg_quality: int,
+) -> list[Path]:
+    expr = "select=" + "+".join(f"eq(n\\,{frame})" for frame in FRAME_STARTS)
+    out_pattern = frame_dir / "frame_%02d.jpg"
+    cmd = [
+        ffmpeg,
+        "-hide_banner",
+        "-loglevel",
+        "error",
+        "-y",
+        "-i",
+        str(mp4_path),
+        "-vf",
+        expr,
+        "-fps_mode",
+        "passthrough",
+        "-q:v",
+        str(jpeg_quality),
+        str(out_pattern),
+    ]
+    subprocess.run(cmd, check=True)
+    frames = sorted(frame_dir.glob("frame_*.jpg"))
+    if len(frames) != SEGMENTS_PER_VIDEO:
+        raise RuntimeError(f"expected {SEGMENTS_PER_VIDEO} frames, got {len(frames)}")
+    return frames
+
+
+def write_npz_bytes(poses: np.ndarray, intrinsics: np.ndarray) -> bytes:
+    buf = io.BytesIO()
+    np.savez_compressed(
+        buf,
+        poses=poses.astype(np.float32, copy=False),
+        intrinsics=intrinsics.astype(np.float32, copy=False),
+    )
+    return buf.getvalue()
+
+
+def process_shard(args: tuple) -> dict:
+    (
+        shard_idx,
+        keys,
+        out_dir_str,
+        jpeg_quality,
+    ) = args
+    out_dir = Path(out_dir_str)
+    shard_dir = out_dir / "shards"
+    parts_dir = out_dir / "manifest_parts"
+    shard_path = shard_dir / f"sekai_game_train_10s_camera_segments-{shard_idx:06d}.tar"
+    part_path = parts_dir / f"part-{shard_idx:06d}.jsonl"
+    tmp_root = out_dir / "tmp"
+    tmp_root.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg = imageio_ffmpeg.get_ffmpeg_exe()
+    camera = np.load(CAMERA_PATH, allow_pickle=False)
+    ids = camera["ids"]
+    ranges = camera["ranges"]
+    poses_all = camera["pose"]
+    intrinsics_all = camera["intrinsics"]
+    id_to_idx = {str(key): i for i, key in enumerate(ids)}
+
+    sample_count = 0
+    image_bytes = 0
+    camera_bytes = 0
+    started = time.time()
+
+    with zipfile.ZipFile(ZIP_PATH) as zf, tarfile.open(shard_path, "w") as tar, part_path.open(
+        "w", encoding="utf-8"
+    ) as manifest:
+        with tempfile.TemporaryDirectory(dir=tmp_root) as tmp:
+            tmp = Path(tmp)
+            for key in keys:
+                if key not in id_to_idx:
+                    raise KeyError(f"{key} missing from camera ids")
+
+                mp4_path = tmp / f"{key}.mp4"
+                frame_dir = tmp / f"{key}_frames"
+                frame_dir.mkdir(exist_ok=True)
+
+                with zf.open(f"{key}.mp4") as src, mp4_path.open("wb") as dst:
+                    shutil.copyfileobj(src, dst, length=1024 * 1024)
+                meta = json.loads(zf.read(f"{key}.json").decode("utf-8"))
+                prompt = meta["prompt"]
+
+                idx = id_to_idx[key]
+                camera_start, camera_len = map(int, ranges[idx])
+                if camera_len < SEGMENTS_PER_VIDEO * SEGMENT_FRAMES:
+                    raise ValueError(f"{key} camera length {camera_len} is shorter than 960")
+                poses = poses_all[camera_start : camera_start + camera_len]
+                intrinsics = intrinsics_all[camera_start : camera_start + camera_len]
+
+                frame_paths = extract_segment_frames(ffmpeg, mp4_path, frame_dir, jpeg_quality)
+                for segment_index, frame_start in enumerate(FRAME_STARTS):
+                    frame_end = frame_start + SEGMENT_FRAMES
+                    sample_id = make_sample_id(key, segment_index, frame_start)
+                    base = f"samples/{sample_id}"
+                    image_data = frame_paths[segment_index].read_bytes()
+                    camera_data = write_npz_bytes(
+                        poses[frame_start:frame_end],
+                        intrinsics[frame_start:frame_end],
+                    )
+                    metadata = {
+                        "id": sample_id,
+                        "source_video_key": key,
+                        "segment_index": segment_index,
+                        "frame_start": frame_start,
+                        "frame_end_exclusive": frame_end,
+                        "num_frames": SEGMENT_FRAMES,
+                        "duration_seconds": SEGMENT_SECONDS,
+                        "fps": FPS,
+                        "input_image": f"{base}.jpg",
+                        "prompt": f"{base}.txt",
+                        "camera": f"{base}.npz",
+                        "pose_shape": [SEGMENT_FRAMES, 4, 4],
+                        "intrinsics_shape": [SEGMENT_FRAMES, 4],
+                        "source_width": meta.get("width"),
+                        "source_height": meta.get("height"),
+                        "source_prompt_sha256": hashlib.sha256(prompt.encode("utf-8")).hexdigest(),
+                        "source_others": meta.get("others", {}),
+                    }
+
+                    add_bytes_to_tar(tar, f"{base}.jpg", image_data)
+                    add_bytes_to_tar(tar, f"{base}.txt", prompt.encode("utf-8"))
+                    add_bytes_to_tar(
+                        tar,
+                        f"{base}.json",
+                        json.dumps(metadata, ensure_ascii=False, sort_keys=True).encode("utf-8"),
+                    )
+                    add_bytes_to_tar(tar, f"{base}.npz", camera_data)
+
+                    manifest_entry = {
+                        **metadata,
+                        "shard": f"shards/{shard_path.name}",
+                        "tar_members": {
+                            "input_image": f"{base}.jpg",
+                            "prompt": f"{base}.txt",
+                            "metadata": f"{base}.json",
+                            "camera": f"{base}.npz",
+                        },
+                    }
+                    manifest.write(json.dumps(manifest_entry, ensure_ascii=False) + "\n")
+                    sample_count += 1
+                    image_bytes += len(image_data)
+                    camera_bytes += len(camera_data)
+
+                mp4_path.unlink(missing_ok=True)
+                shutil.rmtree(frame_dir, ignore_errors=True)
+
+    return {
+        "shard": str(shard_path),
+        "part": str(part_path),
+        "videos": len(keys),
+        "samples": sample_count,
+        "image_bytes": image_bytes,
+        "camera_bytes": camera_bytes,
+        "elapsed_seconds": round(time.time() - started, 3),
+    }
+
+
+def chunks(items: list[str], size: int) -> list[list[str]]:
+    return [items[i : i + size] for i in range(0, len(items), size)]
+
+
+def build_readme(out_dir: Path, stats: dict) -> None:
+    readme = f"""# sekai_game_train_10s_camera_segments
+
+由 `{SRC_DIR}` 整理得到。该版本不包含原始视频；每条样本只包含一张输入图、一段 prompt、以及对应 10 秒的逐帧相机参数。
+
+## 采样规则
+
+- 原始视频：1618 条，每条 961 帧，16 fps。
+- 每条视频按非重叠 10 秒片段采样：每段 160 帧。
+- 每条原视频取 6 段：`[0,160)`, `[160,320)`, `[320,480)`, `[480,640)`, `[640,800)`, `[800,960)`。
+- 每段保存该段第 1 帧作为 `input_image`，保存原视频 prompt，保存该段 160 帧的 `poses` 和 `intrinsics`。
+- 原视频多出的最后第 961 帧不单独成段。
+
+## 数据格式
+
+数据以 WebDataset 风格 tar shards 存在 `shards/` 中。每个样本包含：
+
+- `samples/<id>.jpg`：10 秒片段首帧图像。
+- `samples/<id>.txt`：prompt。
+- `samples/<id>.npz`：包含 `poses (160, 4, 4)` 和 `intrinsics (160, 4)`。
+- `samples/<id>.json`：元数据。
+
+根目录文件：
+
+- `manifest.jsonl`：每条样本一行，记录 shard、tar member、source key、segment index、frame range 等。
+- `stats.json`：整理统计。
+
+## 统计
+
+- source videos: {stats["source_videos"]}
+- segments per video: {stats["segments_per_video"]}
+- total samples: {stats["total_samples"]}
+- segment frames: {stats["segment_frames"]}
+- segment seconds: {stats["segment_seconds"]}
+- fps: {stats["fps"]}
+- shards: {stats["num_shards"]}
+"""
+    (out_dir / "README_zh.md").write_text(readme, encoding="utf-8")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--out-dir",
+        type=Path,
+        default=Path.cwd() / "sekai_game_train_10s_camera_segments",
+    )
+    parser.add_argument("--workers", type=int, default=8)
+    parser.add_argument("--videos-per-shard", type=int, default=100)
+    parser.add_argument("--jpeg-quality", type=int, default=3)
+    parser.add_argument("--max-videos", type=int, default=None)
+    args = parser.parse_args()
+
+    out_dir = args.out_dir.resolve()
+    shard_dir = out_dir / "shards"
+    parts_dir = out_dir / "manifest_parts"
+    for path in [out_dir, shard_dir, parts_dir]:
+        path.mkdir(parents=True, exist_ok=True)
+
+    with zipfile.ZipFile(ZIP_PATH) as zf:
+        keys = sorted(Path(name).stem for name in zf.namelist() if name.endswith(".mp4"))
+    if args.max_videos is not None:
+        keys = keys[: args.max_videos]
+
+    work = [
+        (idx, chunk, str(out_dir), args.jpeg_quality)
+        for idx, chunk in enumerate(chunks(keys, args.videos_per_shard))
+    ]
+
+    print(
+        json.dumps(
+            {
+                "out_dir": str(out_dir),
+                "source_videos": len(keys),
+                "shards": len(work),
+                "workers": args.workers,
+                "segments_per_video": SEGMENTS_PER_VIDEO,
+                "expected_samples": len(keys) * SEGMENTS_PER_VIDEO,
+            },
+            indent=2,
+        ),
+        flush=True,
+    )
+
+    results = []
+    with mp.Pool(processes=args.workers) as pool:
+        for result in pool.imap_unordered(process_shard, work):
+            results.append(result)
+            print(json.dumps(result, ensure_ascii=False), flush=True)
+
+    results.sort(key=lambda item: item["shard"])
+    manifest_path = out_dir / "manifest.jsonl"
+    with manifest_path.open("w", encoding="utf-8") as out:
+        for result in results:
+            out.write(Path(result["part"]).read_text(encoding="utf-8"))
+
+    total_samples = sum(item["samples"] for item in results)
+    stats = {
+        "source_dir": str(SRC_DIR),
+        "source_zip": str(ZIP_PATH),
+        "source_camera_npz": str(CAMERA_PATH),
+        "source_videos": len(keys),
+        "segments_per_video": SEGMENTS_PER_VIDEO,
+        "total_samples": total_samples,
+        "expected_samples": len(keys) * SEGMENTS_PER_VIDEO,
+        "segment_frames": SEGMENT_FRAMES,
+        "segment_seconds": SEGMENT_SECONDS,
+        "fps": FPS,
+        "frame_starts": FRAME_STARTS,
+        "num_shards": len(results),
+        "videos_per_shard": args.videos_per_shard,
+        "jpeg_quality": args.jpeg_quality,
+        "image_bytes": sum(item["image_bytes"] for item in results),
+        "camera_npz_bytes": sum(item["camera_bytes"] for item in results),
+        "shards": results,
+    }
+    (out_dir / "stats.json").write_text(json.dumps(stats, ensure_ascii=False, indent=2), encoding="utf-8")
+    build_readme(out_dir, stats)
+
+    shutil.rmtree(out_dir / "tmp", ignore_errors=True)
+    print(json.dumps({"done": True, "stats": stats}, ensure_ascii=False), flush=True)
+
+
+if __name__ == "__main__":
+    main()

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000000.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38834d9585732ad227471accbe6f8b3e7d462abf170bcbc3b8c6b66ef0706e9b
+size 153313280

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000001.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7d50c676c6d66ff9b45a8c23bc7abda2e1164515338660311e7464518308672
+size 71137280

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000002.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89d6e20c132417a05d2344fdeb2522de0884df3313d6fb252cd27417d4742ddd
+size 105093120

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000003.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:961ec444cee04f56b2c6326e031410b73e095c53be04232c5bfecd45c89a0cbe
+size 83056640

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000004.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:442385f4aa7cc951d545b76f86505285225dd26ee16ab3bd582cd5c7e8677afa
+size 40396800

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b393d53782e496f923617d4ebf0386c8610920ddecc764a1b5dd36ce86b4d8d0
+size 82677760

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000006.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ce3884e4c344ca01b18d446abd3fe78cf066b2f3db89d547d136081a425ce57
+size 77967360

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000007.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e285f89e91aff6e82a6b653abcc2a0c957e952055a681d3fba7ad7282d9bad1c
+size 81479680

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000008.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df19cee402f0e8af08914bdb1070eee77da45b4f0d0dcf985a76a17a4844da2b
+size 177950720

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000009.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eacb862bca25e437cf116b194b76b9a9f7d6c6d9eb650c56a14ae0930410239a
+size 87132160

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000010.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e7504c6a91773b4e282b26e6ff8176da95b97a83de3c8c742898f4de3614cab
+size 45240320

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000011.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfb6bd922c8cca162484130c43c42617ba1726c8637c914bae46928583f46486
+size 75151360

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000012.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35c2568cc6237cf22ee140677d0ac76326e0e166f5060c14699d8bbfda91236a
+size 56248320

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000013.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abefdc77bb96e8d2b908115e65a3bf80e29b08a2cab81e5ab28105c77f8323c4
+size 61347840

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000014.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd3175fab015455b67896a1f9a4fb77f5722a3481af362c8cf08752daf8ad8d1
+size 147374080

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000015.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93c90c0747da0715475065c50ce118c1eab3f1d05bbfeaad99943f5e7cc97ad7
+size 130406400

sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000016.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5ae772e92d1f7b26a917416439d512843c324f4756812ce883a02d554f28d47
+size 7403520

sekai_game_train_10s_camera_segments/stats.json

@@ -0,0 +1,180 @@
+{
+  "source_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/haozhu/data/sekai_game_train_961frames_16fps_ovl640",
+  "source_zip": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/haozhu/data/sekai_game_train_961frames_16fps_ovl640/sekai_game_train_00000000.zip",
+  "source_camera_npz": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/haozhu/data/sekai_game_train_961frames_16fps_ovl640/sekai_game_train_00000000_camera.npz",
+  "source_videos": 1618,
+  "segments_per_video": 6,
+  "total_samples": 9708,
+  "expected_samples": 9708,
+  "segment_frames": 160,
+  "segment_seconds": 10,
+  "fps": 16.0,
+  "frame_starts": [
+    0,
+    160,
+    320,
+    480,
+    640,
+    800
+  ],
+  "num_shards": 17,
+  "videos_per_shard": 100,
+  "jpeg_quality": 3,
+  "image_bytes": 1373623035,
+  "camera_npz_bytes": 55022040,
+  "shards": [
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000000.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000000.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 146797817,
+      "camera_bytes": 3123411,
+      "elapsed_seconds": 834.137
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000001.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000001.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 64199643,
+      "camera_bytes": 3538978,
+      "elapsed_seconds": 565.707
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000002.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000002.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 98398961,
+      "camera_bytes": 3314484,
+      "elapsed_seconds": 709.401
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000003.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000003.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 76561022,
+      "camera_bytes": 3115598,
+      "elapsed_seconds": 730.798
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000004.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000004.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 33448313,
+      "camera_bytes": 3554401,
+      "elapsed_seconds": 532.903
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000005.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000005.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 75804134,
+      "camera_bytes": 3481634,
+      "elapsed_seconds": 841.066
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000006.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000006.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 70806115,
+      "camera_bytes": 3779717,
+      "elapsed_seconds": 677.297
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000007.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000007.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 74697893,
+      "camera_bytes": 3410153,
+      "elapsed_seconds": 602.579
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000008.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000008.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 170912106,
+      "camera_bytes": 3664881,
+      "elapsed_seconds": 807.043
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000009.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000009.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 80149311,
+      "camera_bytes": 3595245,
+      "elapsed_seconds": 689.583
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000010.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000010.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 38515145,
+      "camera_bytes": 3342344,
+      "elapsed_seconds": 521.748
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000011.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000011.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 68126048,
+      "camera_bytes": 3646752,
+      "elapsed_seconds": 604.423
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000012.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000012.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 49621641,
+      "camera_bytes": 3250103,
+      "elapsed_seconds": 560.647
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000013.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000013.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 55188767,
+      "camera_bytes": 2787352,
+      "elapsed_seconds": 565.895
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000014.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000014.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 140847327,
+      "camera_bytes": 3152099,
+      "elapsed_seconds": 609.483
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000015.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000015.jsonl",
+      "videos": 100,
+      "samples": 600,
+      "image_bytes": 123322592,
+      "camera_bytes": 3704417,
+      "elapsed_seconds": 599.8
+    },
+    {
+      "shard": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/shards/sekai_game_train_10s_camera_segments-000016.tar",
+      "part": "/lustre/fsw/portfolios/nvr/projects/nvr_elm_llm/users/shyang/wmdata/sekai_game_train_10s_camera_segments/manifest_parts/part-000016.jsonl",
+      "videos": 18,
+      "samples": 108,
+      "image_bytes": 6226200,
+      "camera_bytes": 560471,
+      "elapsed_seconds": 119.368
+    }
+  ]
+}