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/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 342, 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: Column() changed from object to string in row 0
During handling of the above exception, another exception occurred:
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/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 304, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
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.
VetoWorld occasion log
A daily record of whether a hosted model still behaves the way it did yesterday.
Every row is one channel, one provider, one UTC day: a fixed grid of episodes served against a frozen anchor, with a Fisher test and a verdict. The point is not the models. The point is the serving.
Why this exists
The VetoWorld programme measured one model's rate shift 0.319 between two serving days, on the same pinned world, same prompt, same request form. Batch composition and prefix cache were both ruled out; a deployment change on the provider's side is consistent with it and untestable from outside.
That single observation is expensive to live with. It means any two numbers from a hosted endpoint separated in time may differ because the endpoint changed, and nothing in a normal benchmark run would tell you. So this log exists to make the question answerable rather than arguable: it serves a fixed grid every day and records what came back.
A second use appeared and is now the more important one. The fleet schedules the science. Some probes require the state to be expressed — a comparison that measures the damping of a shift is meaningless on a day with no shift — so the daily reading is what grants permission to spend. A skipped day is a day the programme cannot react, not merely a gap in monitoring.
What a row is
| field | meaning |
|---|---|
date |
UTC serving day |
provider |
who answered, taken from the response, not from the request |
channel |
world + arm, e.g. LAT.A0 |
verdict |
QUIET, EVENT, STALE, NO-ANCHOR |
direction |
up or down — EVENT means MOVED, not FELL |
now |
successes / episodes on the day |
p_epoch |
Fisher vs the frozen epoch anchor |
p_rolling |
Fisher vs the rolling window of recent QUIET days |
status |
OK, PARTIAL, SERVE_FAIL, VERDICT_FAIL, BUDGET_REFUSED |
probe_pin |
the payload digest the day was served under |
Five things to know before using it
1. Levels are never compared across providers. The same open weights on a different serving stack are a different served artifact. Only within-provider deltas and cross-provider coincidence of events are read. A row from one provider and a row from another are not two measurements of one thing.
2. Direction is part of the verdict. Reading EVENT as "fell" inverted a conclusion once in this programme: two co-located worlds rose on the day a third fell, and a one-sided reading sleeps through half the signal.
3. False alarms are expected and pre-counted. At three channels, 365 days and alpha = 0.01, roughly 11 EVENT-days per year arise from noise alone. That figure is printed beside every verdict. An isolated EVENT is not a finding; a coincidence across independent columns, or a persistent run, might be.
4. A missing day is not a quiet day. STALE and SERVE_FAIL are distinct
verdicts and neither means the channel was calm. VERDICT_FAIL exists
specifically so that a day where serving worked and the computation broke can
never be mistaken for QUIET.
5. The pin can be amended, and rows say which pin they ran under. This is a
standing instrument, not a fixed-length experiment; its grid and its gates can
change at a stated boundary. Every row carries probe_pin, and superseded pins
stay recomputable, so a reader can always tell which rules produced a row rather
than assuming today's.
Columns
Started 2026-08-15. Together daily; DeepInfra on Mon/Wed/Fri through the
HuggingFace router, where the response header x-inference-provider decides
whether a cell is kept — a routed request that lands elsewhere produces no
row rather than a mislabelled one.
Fireworks and SambaNova are specified and not built: their keys do not exist. Every column begins with no anchor and earns one, so a column started later is not a column started worse.
Reproducing a verdict
pip install vetoworld
vworld occasion verdict --date YYYY-MM-DD
That recomputes from committed cells, costs nothing and needs no key. The serving verb is separate and is the only one that spends.
Limits worth stating plainly
This log cannot tell you why a channel moved. It observes an endpoint from outside, and the provider's side is not visible: a deployment, a quantisation change, a routing change and a silent model swap are indistinguishable here. What it can do is establish when something changed, and whether it changed on more than one stack at once — which is the difference between "this vendor shipped something" and "the ecosystem moved".
It is also a small instrument. A handful of channels at m = 24 has modest power; it detects large shifts reliably and small ones not at all. The 0.319 that motivated it would be caught easily. A 0.05 drift would not.
- Downloads last month
- 81