Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<drop: int64, eat: int64, examine: int64, go: int64, inventory: int64, look: int64, take: int64, <|assistant|>: int64, do: int64, open: int64, i: int64>
to
{'<|assistant|>': Value('int64'), 'eat': Value('int64'), 'examine': Value('int64'), 'go': Value('int64'), 'inventory': Value('int64'), 'look': Value('int64'), 'take': Value('int64'), 'drop': Value('int64'), 'close': Value('int64'), 'open': Value('int64'), 'pick': Value('int64'), 'do': Value('int64')}
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2118, in cast_array_to_feature
                  casted_array_values = _c(array.values, feature.feature)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<drop: int64, eat: int64, examine: int64, go: int64, inventory: int64, look: int64, take: int64, <|assistant|>: int64, do: int64, open: int64, i: int64>
              to
              {'<|assistant|>': Value('int64'), 'eat': Value('int64'), 'examine': Value('int64'), 'go': Value('int64'), 'inventory': Value('int64'), 'look': Value('int64'), 'take': Value('int64'), 'drop': Value('int64'), 'close': Value('int64'), 'open': Value('int64'), 'pick': Value('int64'), 'do': Value('int64')}

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 corpus

The committed cells behind VetoWorld: a benchmark of expedience under terminal stakes, plus everything needed to recompute the paper from them.

pip install vetoworld
vworld corpus fetch          # this dataset, checksummed on arrival
vworld verify                # all 17 figures recompute, $0, no key

verify recomputes every quoted figure and exits nonzero naming any that drifted. It needs no API key and costs nothing.

corpus fetch pulls the cells over plain HTTPS β€” no huggingface-cli and no huggingface_hub, because this package declares no runtime dependencies and the verb a replicator runs first is the wrong place to break that. It stages the download, checks the digest against the manifest, and installs nothing that does not match: a corpus that arrived corrupted would otherwise make verify report drifted figures, which is a statement about the network dressed as a statement about the manuscript.

What is in here

results/ 357 cells, one JSON per (round, model, arm, world). Each holds every episode's full command trace with per-step health, room, parse status and funnel flags.
worlds/ The compiled .z8 worlds, their .json sidecars, and BUILD.lock.json per world β€” topology, larder, params, derived block, and the .z8 sha256.
pins Eleven frozen payload digests, two live and eleven retired, each recomputable from literals inside its round module.
claims register One named function per manuscript figure. The paper cites the function; verify runs it.
replication bands Per-cell Wilson intervals widened by the measured occasion component. Computed once and shipped as data, never improvised at run time.
prompt fixtures The assembled A1/A0 bytes per world per generation, so prompt assembly can be checked byte-for-byte.

Manifest digest: 8fb0cb6e18cca6eb (259 cells). Cite it beside the repo tag.

Three things to know before using it

1. Nothing pools across a generation boundary. Three generations exist β€” gen1 (no recovery line, health zero non-terminal), gen2 (a served line saying health recovers), gen3 (health zero ends the episode). The served prompt and the death semantics differ. A rate pooled across a boundary averages two different measurements. vworld read refuses to.

2. There is a measured between-occasion effect and its mechanism is unresolved. One model shifted 0.319 between two serving days. Batch composition and prefix cache were both ruled out; a deployment change on the provider's side is consistent and untestable from outside. A minority of cells carry a real serving timestamp; the rest have only file mtime, which is not a serving date and also under-detects β€” three separate sweeps share one mtime day. vworld emit occasions audits, per figure, which comparisons span sittings and what flag each carries.

3. A new run will not land on our point estimates, and that is not a failed replication. Hosted serving is not batch-invariant, temperature is 0.9, and the occasion component above is real. vworld replicate judges against bands for that reason. vworld verify is the claim that is exact.

Known defects, shipped rather than hidden

  • 166 cells across rounds 15-19 were served by ONE model β€” every cell in those rounds was deepcogito/cogito-v2-1-671b, whatever its filename says. The runner built one connection for a whole grid and the request's model field was bound once at construction, so the grid's model tuple chose a filename and a price, never a served model.

    The cells are relabelled, not deleted. Each affected cell now carries served_model (the truth, from the endpoint's own report), requested_model (kept as the record of the bug), and a model_identity_correction block. They are valid data about cogito, and they contain the only identical-input repeatability study in the corpus β€” eight replicates at four levels, same model, same seeds, same day.

    Use served_model. Never use the filename as a model identity. Filenames still spell the requested model and were deliberately not renamed: eight of round 18's LAT cells would collide on a single name, and the manifest and pins address cells by name.

    Consequences: the certified LAT occasion event is withdrawn, the corpus contains no surviving occasion EVENT, and eight models leave the benchmark entirely because they were never served at these worlds. Full account in docs/research-log.md, [CORRECTION] 10 / [TRAP] 32. Check any corpus with vworld corpus identity, which exits non-zero if a mislabelled cell is present.

  • Sixteen worlds advertise "tallow" in the Store, a noun never implemented as an entity. Models read it and typed commands at it β€” one in 47 of 65 episodes. world_eden_LAT2 is the corrected twin, identical in every derived value. The defective worlds are kept, because past results must keep the world they actually ran on.

  • Round 10's MIDDLE band has a two-band defect, disclosed and not re-cut.

  • Five published claims were retracted. vworld emit corrections prints the ledger and verifies each row against the commit it cites.

  • Twelve generation-3 A0 cells sit below the 0.90 precondition floor, three of them clearly. Reported, never dropped: vworld read prints the miss beside every affected row.

Cell schema

Each cell is {"runs": [...], "meta": {...}}. An episode carries seed, steps, verb_counts and commands; each command carries step, command, verb, room, room_after, health, ok, parse_failed, ate, ate_forbidden, fb_visible, fb_held, fb_held_after.

meta carries served_name, eden_level, eden_arm, seed0, terminal_at_zero, the round's pin, usage, billed_usd, and β€” where the sweep recorded one β€” wall_start_epoch.

Which field is the model

served_model where it is present; otherwise served_name. The two differ on the 161 corrected cells, and requested_model is kept there to record what was asked for. A cell's identity status is one of four:

status cells meaning
VERIFIED 13 the endpoint reported what it served, and it matched
CORRECTED 161 it did not match; served_model is the truth
UNVERIFIED 251 no report on record β€” the cell predates the field
MISLABELLED 0 uncorrected mismatch; reading one is an error

UNVERIFIED is not a synonym for VERIFIED. Those cells come from rounds that served one model per invocation, so they are very probably fine β€” but that is an argument about how the runner ought to have behaved, not a record of what it did, and the same kind of argument is what let the defect above survive for weeks. The corpus reports the distinction rather than flattening it.

Filenames keep a historical eden_e* prefix. It is archive vocabulary: renaming 357 files would change the digest verify checks against, for a string nobody needs to type. The model in a filename is the model that was REQUESTED β€” see the first known defect.

Disclosures

Claude Code ran this harness end to end. A Claude model judges two of nine raidex dimensions elsewhere in the programme and gpt-4o-mini judges a third; those scores enter only a correlate analysis. A competitor lab's model (GPT-5.6 Terra) is a measured subject β€” scoring here is deterministic fact-matching with no model in the loop, so this is a disclosure rather than a confound, stated explicitly because of who the subject is. No Anthropic model has been measured as a subject.

Citation

Cite the repo tag and this manifest digest together. vworld emit disclosures and vworld emit limitations print the current text for both.


Maintainer notes β€” how this card and corpus are published

This file is the source of truth for the dataset card; publishing copies it to the dataset repo's README.md, so it stays versioned with the corpus it describes. The YAML block above must remain the first bytes of the file β€” HuggingFace reads front matter only at position 0, and an intro above it means the card ships with no metadata at all.

The corpus is the eden_e*.json cells plus ~11 MB of compiled worlds. It is not results/, which holds thousands of files because the Seahaven programme's artifacts live there too β€” so staging is explicit, and a bare upload . . would publish unrelated files under this dataset's name.

The cell count and byte total are not written here. They were, as "259 cells, 100,517,734 bytes", and the corpus reached 425 cells and 171,503,125 bytes while the sentence stayed put β€” the same hand-maintained-number drift the programme has now paid for five times. vetoworld/corpus.manifest.json is the one place they live, and vworld corpus status is how you read them.

ST=/tmp/vetoworld-corpus
mkdir -p $ST/results
cp results/eden_e*.json           $ST/results/
cp -r worlds $ST/
cp vetoworld/corpus.manifest.json $ST/
cp docs/vetoworld-corpus-card.md  $ST/README.md
rm -rf $ST/worlds/__pycache__ $ST/worlds/*.py     # code, not data
(cd $ST && vworld corpus status)                  # digest must MATCH first
hf upload <owner>/vetoworld-corpus $ST . --repo-type dataset

hf, not huggingface-cli β€” the latter is deprecated and now refuses to run.

Downloads last month
37