The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
schema: string
completed: bool
zip_file: string
zip_bytes: int64
zip_sha256: string
all_manifest_files_checked: int64
vendored_cargo_checksum_files: int64
full_train_unverified: bool
python_executed: bool
all_mapped_batches_accepted: bool
train_roots: int64
immutable_legacy_offset_semantics: bool
files: list<item: struct<file: string, sha256: string, bytes: int64>>
child 0, item: struct<file: string, sha256: string, bytes: int64>
child 0, file: string
child 1, sha256: string
child 2, bytes: int64
completed_epoch: int64
created_utc: string
full1000_epoch_execution_verified: bool
full_generator_speedup_verified: bool
to
{'all_mapped_batches_accepted': Value('bool'), 'created_utc': Value('string'), 'train_roots': Value('int64'), 'schema': Value('string'), 'full1000_epoch_execution_verified': Value('bool'), 'completed_epoch': Value('int64'), 'full_generator_speedup_verified': Value('bool'), 'files': List({'file': Value('string'), 'sha256': Value('string'), 'bytes': Value('int64')}), 'immutable_legacy_offset_semantics': Value('bool')}
because column names don't match
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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 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 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 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
schema: string
completed: bool
zip_file: string
zip_bytes: int64
zip_sha256: string
all_manifest_files_checked: int64
vendored_cargo_checksum_files: int64
full_train_unverified: bool
python_executed: bool
all_mapped_batches_accepted: bool
train_roots: int64
immutable_legacy_offset_semantics: bool
files: list<item: struct<file: string, sha256: string, bytes: int64>>
child 0, item: struct<file: string, sha256: string, bytes: int64>
child 0, file: string
child 1, sha256: string
child 2, bytes: int64
completed_epoch: int64
created_utc: string
full1000_epoch_execution_verified: bool
full_generator_speedup_verified: bool
to
{'all_mapped_batches_accepted': Value('bool'), 'created_utc': Value('string'), 'train_roots': Value('int64'), 'schema': Value('string'), 'full1000_epoch_execution_verified': Value('bool'), 'completed_epoch': Value('int64'), 'full_generator_speedup_verified': Value('bool'), 'files': List({'file': Value('string'), 'sha256': Value('string'), 'bytes': Value('int64')}), 'immutable_legacy_offset_semantics': Value('bool')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Quantum generator
Synthetic full-matrix CNOT synthesis data for the all-to-all unit-cost alphabet.
Current release
full_shared_native_holdout_20261001.sqlite contains one shared holdout store with 1792 dev and 1792 test roots. The matrix widths are 8, 10, 13, 25, 26, 34, 45, 50, 58, 72, 90, 108, 122 and 144. Root families are walk mixture (ordinary/Poisson), guarded exact, basis shear and composition.
Each root includes the full matrix, complete verified reduction certificates, seed, split and lineage, an audit, canonical matrix identity and a content digest. Exact distance is only present when a separate lower-bound certificate supports the complete upper-bound word. Non-exact compiler words are upper bounds.
The eight guarded exact families are bipartite graph, pair, zero excess, forest, moving-root forest, swap forest, saturated frame and one-excess word bracket. Word-bracket certificates do not count amplitudes.
Completed native train epoch0
full_native_train_epoch0_20261002.sqlite contains10000verified globally unique train roots. See TRAIN_EPOCH0_RELEASE.md, the complete readonly audit and actual30000batch reader receipt. All unaugmented modes passed; three2/augmented batches were rejected by both CPU/CUDA full-bank admission, and the1/augmented mode fails at5110. This is fail-closed evidence, not uninterrupted augmented training readiness. No Python was executed. The1000epoch stream and fresh matched generator speedup remain unverified.
Native migration evidence
The native Rust + C++ generator reproduced complete certificate-portfolio records for all 3584 stored roots. Complete exact records and audits matched for all 896 exact roots. The latest full run also matched complete semantic audits for all 3584 roots; only elapsed seconds and solver_seconds are excluded from equality. A separate Rust read-only replay verified all 6272 complete word certificates. Direct Rust-owned CUDA Gaussian words matched canonical CPU words for all 3584 matrices, using preallocated pinned/device buffers and an owned stream. These checks did not execute Python.
Holdout suffix-offset arrays are intentionally empty. Train suffixes follow the first minimum-cost word and the last occurrence of each trajectory state.
Remaining gates
This repository does not yet contain the 1000 × 10000 train stream. The durable pure-native train driver and fixed1000-epoch controller are implemented. Completion of all1000actual train epochs, remote integration and matched full-workload end-to-end timing remain unfinished. Component parity does not establish an end-to-end speedup. The sealed holdout exclusion bank does not establish completion of the train stream.
The reduced smoke-subset priority was superseded by user clarification on 2026-10-01 at 16:31. The current target remains the complete generator with C++/CUDA compute and Rust control.
Format
SQLite roots has matrix identity, split/epoch/position/attempt, family/width/seed, JSON record and audit, suffix offsets and root digest. Packed matrix rows are arbitrary-precision integers, including widths above 64 bits; consumers must not round them through floating-point JSON numbers.
No training implementation or Python runtime is required to inspect the SQLite artifact using a native SQLite client. The dataset viewer does not directly decode this SQLite release.
Verified full holdout exclusion bank
The 256 sha256_XX.bin shards contain 126631457 unique full 32-byte canonical operator SHA-256 digests, partitioned by the first byte and sorted lexicographically. bank_manifest.json records their counts, byte sizes and SHA-256 values. Independent verification checked all 126631688 parent and primitive-neighbor insertions from all 25096 finite symmetry views, plus every shard hash, count, partition and strict ordering. No approximate filter or Python execution was used.
The verification receipt is full_compact_bank_verification_20261002.json. Bind the bank manifest SHA-256 db9741b1971e96011673d93605f9853e52f48695dca4fd905fadb10424bfaa6e before train admission. Hash collisions conservatively over-filter; root uniqueness still requires the complete matrix identity.
Native train suffix materialization, inverse/transpose/wire-reversal augmentations, exact-label removal after augmentation, and whole-batch admission before returning any record are implemented. All 27 native Rust tests passed, including compiler-failure incumbent retention. These component checks do not establish completion of an actual train epoch.
Full CUDA admission and native recovery source
cuda_full_finite_views_hash_parity_20261002.json proves exact CPU/CUDA equality of all126631688 complete parent/primitive SHA256 digests from all25096 finite views and14 widths. cuda_full_bank_admission_parity_20261002.json binds that pinned digest-parity receipt and separately verifies the complete GPU-loaded bank rejects all25096 views and matches56 train admission cases. These are complete holdout/bank checks, not completion of the train stream.
native_full_train_recovery_source_20261002.zip adds Rust-owned full-bank CUDA admission, immutable SQLite train prefixes, whole actual batch admission, a fixed1000-epoch controller, exclusive writer ownership and bounded atomic-heartbeat retry on Windows sharing conflicts. Its2022 manifest files and27 vendor checksum manifests were independently verified inside the ZIP. See NATIVE_RECOVERY_PACKAGE.md and native_full_train_recovery_archive_verification_20261002.json.
The source snapshot's37 regular library tests, two controller tests and two heartbeat tests passed. Reader validation reuse preserves complete outputs and admission matrices in252 differential cases, including malformed-certificate and blocked-final-state refusal. The unchanged CUDA kernel also passed70 source suffix fixtures acrossall14 widths. The controller has not executedall1000 epochs. The current train epoch resumed a verified4891-root prefix; no completed10000-root epoch or matched end-to-end speedup is claimed by this source release.
- Downloads last month
- 1,485