Datasets:
The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
RefVideo6M β style v2v quadruples (src video + style ref image + styled GT + prompt)
Filtered subset of RefVideo6M.
Only samples that ship all four assets are kept, i.e. the ones where a style
reference image exists β those are the instruction-following / reference-conditioned
V2V training quadruples. Samples without _img_ref.jpeg (text-only style prompts)
are dropped.
| samples | 79,390 |
| unpacked size | 353.2 GB |
| distinct styles | 305 |
| src/tgt duration | 8.1 s, 10 fps, 81 frames |
| resolutions | 1248x720 / 720x1248 (src == tgt, per sample) |
| ref image size | variable (independent style image, not the first frame) |
| shards | 207 x ~2 GB uncompressed tar |
Layout
packs/pack-XXXX-NNNN.tar # uncompressed (STORED) tar, ~2 GB each
index.jsonl # one row per sample: uid, key, pack, prompt, style, sizes
tarindex.jsonl # per pack: every member's byte offset + size
packs.jsonl # per-pack: bytes, n_samples, first/last uid
stats.json # totals
tarindex.jsonl is the important one: because the tars are uncompressed, a single
sample can be fetched with HTTP range requests β no 2 GB download. See below.
Each tar holds, per sample (flat member names, key = 32-hex uid):
<key>_src_video.mp4 # source video
<key>_tgt_video.mp4 # styled ground truth
<key>_img_ref.jpeg # style reference image
<key>_prompt.txt # JSON: instructions[], src_video_path, ...
Tars are stored uncompressed on purpose: mp4/jpeg are already compressed, the
bytes stay identical to upstream, and the archive stays seekable so you can pull a
single sample without unpacking a whole shard.
Usage
Random sample access (reads only the tar index + the 4 members of the sample):
import json, tarfile
from huggingface_hub import hf_hub_download
REPO = "yeccanbebetter/RefVideo6M_style_v2v_quad"
rows = [json.loads(l) for l in open(hf_hub_download(REPO, "index.jsonl", repo_type="dataset"))]
row = rows[0]
path = hf_hub_download(REPO, f"packs/{row['pack']}", repo_type="dataset")
with tarfile.open(path) as tf:
src = tf.extractfile(row['key'] + "_src_video.mp4").read()
tgt = tf.extractfile(row['key'] + "_tgt_video.mp4").read()
ref = tf.extractfile(row['key'] + "_img_ref.jpeg").read()
prompt = json.loads(tf.extractfile(row['key'] + "_prompt.txt").read())
Single-sample fetch without downloading the shard (uses tarindex.jsonl):
import json, requests
from huggingface_hub import hf_hub_download, hf_hub_url
REPO = "yeccanbebetter/RefVideo6M_style_v2v_quad"
rows = [json.loads(l) for l in open(hf_hub_download(REPO, "index.jsonl", repo_type="dataset"))]
taridx = {}
for l in open(hf_hub_download(REPO, "tarindex.jsonl", repo_type="dataset")):
d = json.loads(l)
for name, (off, size) in d["members"].items():
taridx[name] = (d["pack"], off, size)
row = rows[0]
url = hf_hub_url(REPO, "packs/" + row["pack"], repo_type="dataset")
def get(name):
pack, off, size = taridx[name]
r = requests.get(url, headers={"Range": f"bytes={off}-{off+size-1}"})
return r.content
src = get(row["key"] + "_src_video.mp4")
ref = get(row["key"] + "_img_ref.jpeg")
WebDataset-style streaming (shards are already in the right shape):
import webdataset as wds
ds = wds.WebDataset("hf://datasets/" + REPO + "/packs/pack-{0000..0207}-{0000..0004}.tar")
Filtering already applied (see the upstream build_triples.py): all four files
present, prompt parses, moov probe gives w/h/dur, |dur_src - dur_tgt| <= 0.15 s,
src/tgt same resolution, ref decodes with PIL, prompt paths match filenames,
src != tgt byte-wise, plus a 3-frame visual sanity pass.
- Downloads last month
- 161