The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: JSON parse error: Invalid value. in row 0
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 324, in _generate_tables
df = pandas_read_json(f)
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
return pd.read_json(path_or_buf, **kwargs)
~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 815, in read_json
return json_reader.read()
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1014, in read
obj = self._get_object_parser(self.data)
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1040, in _get_object_parser
obj = FrameParser(json, **kwargs).parse()
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1176, in parse
self._parse()
~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1392, in _parse
ujson_loads(json, precise_float=self.precise_float), dtype=None
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: Unexpected character found when decoding 'Infinity'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4523, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2768, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2972, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2483, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 327, in _generate_tables
raise e
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Invalid value. in row 0Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Spark-H3 vs regular H3: speed, movement and cost
Public evidence for the Agentic Film Lab eight-scene RunPod Serverless experiment, generated October 13, 2026. Compare Dense H3, Spark at 20% attention density with 4 initial dense evaluations, and Spark with 0. Every lane still performs 20 actual model evaluations. Worker compile/cache warm-ups are a separate concept.
Measured results
| Setting | Whole generation, seconds | Sampling, seconds | Whole-generation speedup | Generation-only estimated USD |
|---|---|---|---|---|
| Dense | 719.80 | 621.64 | 1.00x | 0.70 |
| Spark /4 | 393.36 | 314.10 | 1.83x | 0.38 |
| Spark /0 | 339.76 | 261.99 | 2.12x | 0.33 |
These are equal-scene means of scene medians. S01/S07 have 3 measured timing trials per lane; the other scenes have 1. Estimated prices use the recorded 3.49 USD/worker-hour quote and exclude startup, idle time and storage. They are not per-clip invoices. Sampling speedups were 1.98x and 2.37x; queues and cold starts are outside both timed comparisons.
The actual posted completed-run allocation was $37.77: $37.47873872128548 endpoint plus $0.2916666679 for the entire shared 300 GB volume during the run window. The separately dated snapshot through 12:00 UTC on October 13 was $37.80; retained storage continues. Do not add generation-only estimates to these charges.
What is included
Download the complete evidence archive. Its contents and hashes are in the file manifest and SHA256SUMS.
- All 48 original MP4s, unchanged:24 core comparisons,12 additional measured timing repeats,6 cold-pilot clips,6 excluded worker compile/cache warm-ups.
- Actual frozen prompts, seeds, ComfyUI workflows and four appearance-reference JPEGs used for Umbrella Girl and Mountain Man.
- Clip/stage/model-evaluation receipts, runtime and checkpoint locks, source revisions/checksums, worker and provider logs, failures/recovery records and accounting.
- Posted scoped billing snapshots and reconciliation, separated from quote-based per-clip estimates.
- Scene observations, source-linked narration claims, final edit coverage, captions, chapter timestamps and delivery QA.
The archive mirrors episode-relative paths. The primary clips are under serverless/generation-v003/campaign-runs/20261013T060254Z/raw-clips/. Prefixes M-/C-/W- distinguish measured, cold pilot and excluded worker warm-up outputs; use clip-ledger.csv for exact phases. Prompts with the v003 suffix and the frozen generation-v003/manifest.json are the actual test authority; earlier files remain historical.
Method and limits
Two character scenarios each for Umbrella Girl and Mountain Man use Ref2VA; four new subjects use the FL2VA checkpoint in text-only mode without keyframes. Output 1344x768,243 frames at 24 fps = 10.125 seconds. Same prompt/seed/settings within each scene; pruned INT8 convrot H3 weights and NVFP4 AWQ text encoder. RTX PRO 6000 Blackwell Server Edition 96 GB GPUs, same exact SKU but separate physical workers. Six workers ran the main campaign within a ceiling of 10. Spark density 0.2, initial dense evaluations 4 or 0, dense_layers 0.
The creator's full-set impression was that Spark / 4 looked identical in general and Spark / 0 almost identical, with creative differences sometimes preferred. This is personal viewing judgment, not pixel equality or a statistical quality study. Agent visual review sampled playback around 1 Hz and selected exact frames; it does not certify every 24 fps artifact. No aesthetic video rerolls. Initialization, receipt/export failures and recovery without regenerating existing video are preserved.
The narrated episode scales footage to 4K without crop or retiming. Replays and audio lane selection are labeled. Native audio is level-adjusted and ducked beneath narration; raw MP4s here retain their original audio. No final narrated master, private voice-reference recording or narration stems are included.
Sources and licensing boundary
- Spark release/card: https://huggingface.co/Aazeus/Spark-H3
- Spark code tested: https://github.com/zechengtang/Spark-H3/tree/1bb4cdb22648314901e4cbe82c3b414aa1f2e769
- Official H3: https://huggingface.co/MiniMaxAI/MiniMax-H3
- Exact Comfy-Org weight revision: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/e5eb578a89295337b8ff433a035929ce0279e0b6
- ComfyUI: https://github.com/Comfy-Org/ComfyUI
- RunPod Serverless: https://docs.runpod.io/serverless/overview
Spark is an attention backend, not a LoRA or a newly trained checkpoint. Spark code has its own Apache 2.0 license; H3 weights have a separate upstream license. This mixed evidence repository uses the other license marker and does not grant a blanket Apache license over every asset. No weights, wheels or upstream library bundles are redistributed. Consult each upstream project for its terms.
Publication redactions
Credentials, account identity/contact fields, local usernames and private voice-reference fields are removed; infrastructure identifiers use stable aliases so worker variation remains traceable. Redundant encoded download chunks are omitted because their decoded MP4s and receipts are included. Private deployment environment dumps, unneeded upstream documentation copies, binary bundles and edit caches are excluded. Selected archived upstream source files are retained with original revision/hash receipts and their licenses under LICENSES/. Source-record and published-file hashes identify transformations. Logs are retained as recorded with these redactions; no complete provider backfill is claimed.
- Downloads last month
- 65