yslan/holo4d_dataset_caption

Holo4D Gemini caption annotations

This Hugging Face dataset contains semantic caption annotations for Holo4D training sources. It contains annotation JSONL files and provenance metadata; it does not contain the original RGB frames, depth maps, camera parameters, or 3D tracks. Those modalities remain in the Holo4D/unified dataloader data roots.

Syn4D v1 global captions

The Syn4D v1 global export is under syn4d_v1/.

File	Use	Records
global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64_success.jsonl	Recommended clean file; successful records only	4,515
global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64.jsonl	Raw resumable output; includes failed attempts for provenance	4,519
README.md	Syn4D v1-specific notes	—

The clean file currently contains 4,515 successful global captions. The raw file contains four error records. Two source sequences are intentionally not represented by a successful caption because the source is unusable:

• syn4d_v1_stride1/hallway/seq_000012_0: required PNG frames are missing.
• syn4d_v1_stride1/arena/seq_000012_6: Gemini input was blocked by the provider's prohibited-content filter.

Four error lines are retained in the raw file for auditability. One of them, winter/seq_000017_0, also has a later successful record, so it is already present in the clean export. Three unique sequences remain retryable: lab_bald/seq_000010_1, planet_bald/seq_000015_1, and winter_group_static/seq_000013_3. Consumers should use the success-only file unless they explicitly need retry diagnostics.

What “global” means

There is one record per source sequence, rather than one record per 64-frame training window. The complete sequence is represented by at most 64 frames sampled uniformly over its full temporal extent. Therefore:

• annotation.window.num_frames = 64 means 64 sampled observations;
• it does not mean a contiguous 64-raw-frame training clip;
• the exact raw frame indices are in metadata.sampled_frames;
• the full-sequence observation is marked by metadata.observation_scope = "full_sequence";
• annotation.sample_id ends in _global.

The source data uses 30 FPS. The global caption is optional sequence-level context. For VDM training, use the global vdm_prompt as a sequence caption when sampling a short training clip from that same source sequence. The separate window-caption export, when available, is more temporally local and should be preferred when a training sample has an aligned window caption.

Syn4D v2 captions

Syn4D v2 has two local-window exports and one independent global export under syn4d_v2/:

File	Use	Records
captions_gemini-3_8-flash_high_holo4d_semantic_v3_1_raw96_success.jsonl	Recommended clean raw-frame local captions	10,127
captions_gemini-3_8-flash_high_holo4d_semantic_v3_1_raw96.jsonl	Raw resumable local output, including historical failures	10,141
global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64_success.jsonl	Recommended clean one-per-sequence global captions	4,799
global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64.jsonl	Raw resumable global output, including historical failures	4,804
manifest_raw96.json	Local-window manifest for the clean raw96 export	10,127
global_manifest_uniform64.jsonl	Manifest for clean global exports	4,799

The older captions_gemini-3_8-flash_high_holo4d_semantic_v3_1_n064.jsonl file remains available for provenance, but it used the earlier sampled-frame configuration and is not the canonical raw-frame-aligned local export.

Syn4D v2 local raw96 contract

The recommended local captions use:

• 96 consecutive raw frames per Gemini request;
• sample_stride=1;
• 96-frame window starts, with an end-aligned final window when required;
• 30 FPS source timing;
• no camera trajectory in the semantic caption.

The local caption describes the 96-frame semantic block. A 49- or 56-frame VDM crop may be sampled inside that block, but the dataloader should keep the crop within the selected block if it requires strict temporal alignment. The clean raw96 export covers 10,127 of 10,128 planned windows. One window is absent because Gemini rejected its input under the provider policy filter.

The global export is independent of the local windows: it uses at most 64 frames sampled uniformly across each complete sequence. It has 4,799 clean records for 4,800 planned sequences; one sequence remains absent because of the same provider policy filter. Use the _success.jsonl files for training.

Example:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yslan/holo4d_dataset_caption",
    repo_type="dataset",
    filename=(
        "syn4d_v2/"
        "captions_gemini-3_8-flash_high_holo4d_semantic_v3_1_raw96_success.jsonl"
    ),
)
with open(path, encoding="utf-8") as f:
    record = json.loads(next(line for line in f if line.strip()))
prompt = record["vdm_prompt"]
video_id = record["metadata"]["video_id"]

JSONL record format

Each non-empty line is one independent JSON object. The recommended text field for conditioning is:

{
  "vdm_prompt": "A foggy ruins is shown ...",
  "vdm_prompt_variant": "holo4d_semantic_text_v3",
  "annotation": {
    "schema_version": "holo4d_semantic_v3",
    "sample_id": "syn4d_v1_stride1/arena/seq_000003_7_global",
    "source": {
      "video_path": "/vgg/chenzhao/syn4d_v1_stride_1/arena/exr_layers/image/seq_000003_7",
      "dataset": "syn4d_v1_stride1"
    },
    "window": {
      "start_frame": 0,
      "end_frame": 476,
      "num_frames": 64,
      "sampled_fps": 3.9705882353,
      "start_time_sec": 0.0,
      "end_time_sec": 15.8666666667,
      "duration_sec": 15.8666666667
    },
    "scene": {},
    "subjects": [],
    "objects": [],
    "relations": [],
    "timeline": [],
    "visible_text": [],
    "human_readable_summary": "",
    "uncertainty": []
  },
  "metadata": {
    "annotation_model": "gemini-3.8-flash",
    "schema_version": "holo4d_semantic_v3",
    "video_id": "syn4d_v1_stride1/arena/seq_000003_7",
    "scene_name": "arena",
    "sequence_name": "seq_000003_7",
    "raw_fps": 30.0,
    "sample_stride": 1,
    "sampled_frames": [0, 8, 15, 23],
    "observation_scope": "full_sequence",
    "audio_available": false
  }
}

The real records contain the complete lists. The semantic fields mean:

• scene: environment, setting, weather, lighting, ground, structures, and other scene-level context;
• subjects: people, figures, animals, or other salient entities;
• objects: visible non-subject objects and attributes;
• relations: spatial or interaction relations among entities;
• timeline: temporally ordered actions or changes observed in the sequence;
• visible_text: text visible in the sampled frames;
• human_readable_summary: concise natural-language summary;
• uncertainty: observations that should not be treated as certain.

The structured annotation is the semantic source of truth. vdm_prompt is the compiled text conditioning string. Geometry, camera motion, depth, and 3D tracks are not inferred from this caption; they must be loaded from the corresponding source dataset by the unified dataloader.

Download and read from the Hub

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yslan/holo4d_dataset_caption",
    repo_type="dataset",
    filename=(
        "syn4d_v1/"
        "global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64_success.jsonl"
    ),
)

captions = {}
with open(path, "r", encoding="utf-8") as f:
    for line in f:
        record = json.loads(line)
        video_id = record["metadata"]["video_id"]
        captions[video_id] = record

record = captions["syn4d_v1_stride1/arena/seq_000003_7"]
prompt = record["vdm_prompt"]
structured = record["annotation"]

For streaming without downloading the whole file:

import json
from huggingface_hub import hf_hub_url
import requests

url = hf_hub_url(
    repo_id="yslan/holo4d_dataset_caption",
    repo_type="dataset",
    filename=(
        "syn4d_v1/"
        "global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64_success.jsonl"
    ),
)
with requests.get(url, stream=True, timeout=60) as response:
    response.raise_for_status()
    for raw_line in response.iter_lines(decode_unicode=True):
        if raw_line:
            record = json.loads(raw_line)
            # metadata.video_id identifies the source sequence.
            # vdm_prompt is the text condition.

Joining to the local Holo4D data

The paths stored in source.video_path and metadata.source_path are the original machine-local paths and may not exist on another cluster. Use the portable metadata.video_id as the join key:

syn4d_v1_stride1/<scene>/<sequence>

For example:

video_id = syn4d_v1_stride1/arena/seq_000003_7
scene    = arena
sequence = seq_000003_7

The unified dataloader should resolve that key to the local Syn4D v1 RGB, camera, depth, and 3D-track roots. Do not use the absolute /vgg/chenzhao/... path as the cross-machine identifier.

Window captions and other exports

This repository also contains Syn4D v2 captions. Their JSONL records use the same holo4d_semantic_v3 semantic structure, but their sample_id, sampling metadata, and dataset-specific source layout can differ. Always inspect metadata.video_id, metadata.sampled_frames, and metadata.observation_scope rather than assuming that every record is global.

For a local Syn4D v1 window caption, the intended configuration is 64 raw frames with sample_stride=1 and a 32-frame window stride. A window record has an ID of the form:

syn4d_v1_stride1/<scene>/<sequence>_f<start>_f<end>_n064

Global records intentionally do not use that window ID because they describe the complete sequence.

For Syn4D v2 raw96, local records use clip_length=96 and sample_stride=1; their sample_id ends in _n096. The full-sequence global records end in _global and should not be confused with local windows.

Provenance

• Model: gemini-3.8-flash
• Semantic schema: holo4d_semantic_v3
• Global sampler: uniform full-sequence sampling, at most 64 frames
• Source FPS: 30 FPS
• Audio: not used (audio_available=false)
• Source dataset: Syn4D v1 at /vgg/chenzhao/syn4d_v1_stride_1

The raw JSONL is append-only output from a resumable captioning job. It is kept for auditability; use the _success.jsonl export for normal training and evaluation.

Local caption inventory and readiness

This is the current local audit of caption sources and outputs. Counts refer to JSONL records unless a manifest or an official metadata file is named.

Dataset	Caption source and local location	Current coverage	Use now / remaining work
Syn4D v1	Gemini global: /vgg/yushi/dataset/caption_output/gemini_holo4d/syn4d_v1/	4,515 clean global records; 4,519 raw lines	Ready with global captions. Three unique retryable sequences remain; two source sequences are unusable. Local window output is mixed historical data and is not the canonical entry point.
Syn4D v2	Hugging Face syn4d_v2/ raw96 and uniform64 exports	10,127 clean local windows; 4,799 clean global records	Ready. Use the _success.jsonl files. One local window and one global sequence remain unavailable because of provider policy blocks. The older n064 file is retained for provenance.
DL3DV	Gemini local/global under /vgg/yushi/dataset/caption_output/gemini_holo4d/dl3dv/	Global: 4,762 successes + 3 errors. Window manifest: 37,917 records; aggregate JSONL also contains malformed historical lines	Global incomplete after adding 7K. Rerun global to cover all 1K/2K/3K/5K/6K/7K. A new optional 128/128 local runner is provided but not launched.
MVS-Synth	Gemini windows: /vgg/yushi/dataset/caption_output/gemini_holo4d/mvs_synth/	360 successful windows, 52 failed attempts, 120 sequences	Existing local windows are usable where present; global export is not currently present.
TartanAir	Gemini windows: /vgg/yushi/dataset/caption_output/gemini_holo4d/tartanair/	9,366 successful windows, 18 failed attempts, 369 sequences	Existing local windows are mostly usable; global export is not currently present.
ReCamMaster	Gemini windows: /vgg/yushi/dataset/caption_output/gemini_holo4d/recammaster/	109,717 successful windows, 212 failed attempts	The source clips are short enough that one global caption per sequence is sufficient; current global export still needs to be generated.
Kubric49	Gemini n049: /vgg/yushi/dataset/caption_output/gemini_holo4d/kubric49/	11,269 successful records, 3 failed attempts	Already sufficient for 49-frame training. Do not generate another global caption. Native Kubric caption files remain available too.
Kubric81	Gemini n064: /vgg/yushi/dataset/caption_output/gemini_holo4d/kubric81/	7,118 successful records, 93 failed attempts	Already sufficient for 81-frame training. Do not generate another global caption.
OpenVidHD	Official CSV: /vgg/yushi/dataset/OpenVidHD_60k/annotations/local_openvidhd_63337.csv	63,328 caption rows for 63,337 local MP4s	Ready with official captions. Nine local videos have no matching upstream caption row. Use the OpenVid loader/CSV, not Gemini captions.
OmniWorld-Game	Official chunk captions: /vgg/yushi/dataset/OmniWorld_Game_60k/extracted/annotations/OmniWorld-Game/*/text/*.json	Text chunks found for 438 of 479 extracted UIDs	Ready with official frame-range captions. Each JSON has Short_Caption, PC_Caption, Background_Caption, Camera_Caption, Video_Caption, and Key_Tags; this is intentionally chunk-level rather than one forced trajectory caption.

High-resolution local-caption policy

For the high-resolution datasets that still benefit from local captions, DL3DV and Syn4D v1 use the new optional preset:

caption block: 128 contiguous raw frames
block stride: 128 raw frames
training crop: 49 or 64 frames sampled entirely inside one block

This avoids mechanically captioning every 49/81-frame training crop. A 400-frame sequence produces approximately four semantic local blocks. The training dataloader must first choose a caption block and then constrain the random crop start to:

[block_start, block_end - crop_length + 1]

so a 49/64-frame crop never crosses into a different caption block. The global caption remains the sequence-level context and is not a replacement for this alignment rule when local captions are used.

The resumable scripts are:

cd /vgg/yushi/Repo/gemini_captioning
bash examples/caption_dl3dv_local128_resume.sh
bash examples/caption_syn4d_v1_local128_resume.sh

They write new n128 outputs and preserve all existing n064 files. These scripts are not an instruction to rerun Kubric or OpenVid.

Track4D-360 compatibility note

The original Track4D-360 training configs are under /vgg/yushi/Repo/track4d-360/config/track4d_360/. The main data-role configs reference Syn4D, Kubric49/81, DL3DV, MVS-Synth, ReCamMaster, and OMNI-DROID. They expect older cache layouts such as Syn4D {scene}_per_video_chunks.json, or Cut3R-style captions_all.json plus T5 embedding caches. The Gemini JSONL files above are the canonical annotations, but they are not automatically the old Track4D-360 cache format. A conversion layer is still required before using them in an unmodified Track4D-360 job.

OpenVidHD, OmniWorld-Game, MolmoMotion, EgoDex, DROID, and Xperience are not part of the original Track4D-360 training declarations. They use their own caption/metadata loaders and should enter the shared unified_4d_dataloader through dataset-specific adapters.