The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
analysis_population: string
inputs: list<item: struct<model_id: string, path: string, source_sha256: string>>
child 0, item: struct<model_id: string, path: string, source_sha256: string>
child 0, model_id: string
child 1, path: string
child 2, source_sha256: string
protocol_sha256: string
rows: int64
schema_version: int64
status_counts: struct<completed: int64>
child 0, completed: int64
release_label: string
entries: list<item: struct<compression: string, media_type: string, packaged_bytes: int64, packaged_path: str (... 171 chars omitted)
child 0, item: struct<compression: string, media_type: string, packaged_bytes: int64, packaged_path: string, packag (... 159 chars omitted)
child 0, compression: string
child 1, media_type: string
child 2, packaged_bytes: int64
child 3, packaged_path: string
child 4, packaged_sha256: string
child 5, sanitized_uncompressed_bytes: int64
child 6, sanitized_uncompressed_sha256: string
child 7, source_bytes: int64
child 8, source_path: string
child 9, source_sha256: string
to
{'entries': List({'compression': Value('string'), 'media_type': Value('string'), 'packaged_bytes': Value('int64'), 'packaged_path': Value('string'), 'packaged_sha256': Value('string'), 'sanitized_uncompressed_bytes': Value('int64'), 'sanitized_uncompressed_sha256': Value('string'), 'source_bytes': Value('int64'), 'source_path': Value('string'), 'source_sha256': Value('string')}), 'protocol_sha256': Value('string'), 'release_label': Value('string'), 'schema_version': Value('int64')}
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 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 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
analysis_population: string
inputs: list<item: struct<model_id: string, path: string, source_sha256: string>>
child 0, item: struct<model_id: string, path: string, source_sha256: string>
child 0, model_id: string
child 1, path: string
child 2, source_sha256: string
protocol_sha256: string
rows: int64
schema_version: int64
status_counts: struct<completed: int64>
child 0, completed: int64
release_label: string
entries: list<item: struct<compression: string, media_type: string, packaged_bytes: int64, packaged_path: str (... 171 chars omitted)
child 0, item: struct<compression: string, media_type: string, packaged_bytes: int64, packaged_path: string, packag (... 159 chars omitted)
child 0, compression: string
child 1, media_type: string
child 2, packaged_bytes: int64
child 3, packaged_path: string
child 4, packaged_sha256: string
child 5, sanitized_uncompressed_bytes: int64
child 6, sanitized_uncompressed_sha256: string
child 7, source_bytes: int64
child 8, source_path: string
child 9, source_sha256: string
to
{'entries': List({'compression': Value('string'), 'media_type': Value('string'), 'packaged_bytes': Value('int64'), 'packaged_path': Value('string'), 'packaged_sha256': Value('string'), 'sanitized_uncompressed_bytes': Value('int64'), 'sanitized_uncompressed_sha256': Value('string'), 'source_bytes': Value('int64'), 'source_path': Value('string'), 'source_sha256': Value('string')}), 'protocol_sha256': Value('string'), 'release_label': Value('string'), 'schema_version': Value('int64')}
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.
ModelSMC-PBE research artifacts
This dataset is the public data-and-paper companion for Deduce, Propose, Correct: Probability-Accountable LLM Shortlists for Typed Program Synthesis by Tri Nguyen and Thanh-Dat Nguyen. The verified manuscript is 16 pages.
- Dataset: hackerprofile1/modelsmc-pbe-research
- Implementation source and tests: GitHub commit 50369a6b0e264eda9cb6e1aab45949cb3be11cb6
- Manuscript:
paper/main.pdf
The archive preserves successful, failed, diagnostic, superseded, and integrity-limited records separately. A later pass never rewrites an earlier failed gate.
Main evidence
Fresh-blind r5 bounded discovery
On 12 freshly generated finite-domain Filter-then-Map tasks, the complete
LLM-shortlist SMC system found an exact program on 10 tasks versus 2 for matched
grammar-random acquisition. All eight discordant pairs favored the LLM-guided
system (p = 1/256, one-sided exact sign test). This is a bounded system-level
discovery result, not an isolated LLM effect, a speed claim, or broad
generalization evidence.
Calibration chronology
The provider-free calibration evidence must be read in this order:
- Calibration V1 failed. At 256 particles, exact-program-mass RMSE was 0.2395265 and bias was +0.1130184. Exact-program discovery in every primary run did not change the failed gate.
- Terminal diagnostic V2 was diagnostic only. On the same developmental supports, an exhaustive positive control supported poor terminal proposal--target overlap as a material mechanism. It did not rescue V1 or validate the four-stage recurrence.
- Fresh calibration V2 failed. On 20 newly generated tasks, 256-particle RMSE was 0.2612355, bias was -0.0748415, and the task-first bootstrap upper-95 RMSE was 0.3767653. No difficult task was replaced.
- The factorized V3 diagnostic reused the V2 tasks. Its 256-particle RMSE was 0.0221995, but the method was designed after the V2 failure and reused the same tasks and seeds. It was mechanism evidence, not confirmation.
- Fresh V3 R2 passed its narrow frozen gate. A new secret generated 32 independent tasks. At 256 particles, exact-mass RMSE was 0.0398662, bias was -0.0109472, the bootstrap upper-95 RMSE was 0.0494323, and the central 90% bias interval was [-0.0149663, -0.00763969]. Point RMSE decreased to 0.0304513 and 0.0258555 at 512 and 1,024 particles.
Fresh V3 R2 is a provider-free confirmation for terminal exact-program target mass on the declared singleton-complete finite synthetic law. It does not calibrate an LLM posterior, LLM mode banks, the four-stage SMC recurrence, the full PBE system, a large DSL, or external tasks. Target-mean-loss was not the primary endpoint and remained much less accurate (256-particle RMSE 0.2991), so the pass does not extend to that functional.
ExeDec V2 released-data debug
ExeDec V2 observed private-debug-probe exactness in 17/32 paired seed blocks for LLM-SMC versus 3/32 for grammar-random on four public released targets. The strict Filter-then-Map adapter is unofficial, repeated seeds are nested within only four potentially contaminated targets, finite probes are not semantic equivalence, and nonexclusive staging limits integrity. The result is descriptive debug evidence, not a confirmatory ExeDec benchmark.
Artifact locations
- r5 evidence:
artifacts/blind-filter-map-confirmation-v3-r5/ - Calibration V1:
artifacts/particle-calibration-v1/ - Terminal diagnostic V2:
artifacts/particle-calibration-terminal-diagnostic-v2/ - Failed fresh V2:
calibration-v3-r2/artifacts/calibrated-program-inference-v2-fresh/ - Reused-task factorized V3 diagnostic:
calibration-v3-r2/artifacts/calibrated-program-inference-v3-factorized-diagnostic/ - Fresh V3 R2 protocol: frozen source record
- Fresh V3 R2 analysis:
calibration-v3-r2/artifacts/calibrated-program-inference-v3-fresh/analysis.json - Fresh V3 R2 deterministic replay receipt:
calibration-v3-r2/artifacts/calibrated-program-inference-v3-fresh-replay-verification.json - Fresh V3 R2 unblind verification:
calibration-v3-r2/artifacts/calibrated-program-inference-v3-fresh-unblind-verification.json - Fresh V3 R2 release checksums:
calibration-v3-r2/SHA256SUMS - ExeDec V2 canonical study:
artifacts/exedec-deepcoder-ho-debug-benchmark-v2/ - Preserved failed ExeDec operations transfer:
artifacts/exedec-deepcoder-ho-debug-benchmark-v2-operations-evidence-failed-v1/
The Fresh V3 R2 protocol, method seal, custody seal, and deterministic replay
receipt have SHA-256 values
36f4a6c5930c2da2c89c1515d44267b6cc9bab94e7a000352f4343fd97fc97d6,
788a44cc32ed617aa853a9e7ac97ab2236844382825149eba0073cc6d0a31dc4,
32631befddaecd28aa4008c73232475b3676b26a7e1e0fe457dac6995ab0932f,
and c7b08c0385867436e3e2d7a08cfe5a506ed8dfaea702defcabeff1e88b1859a7,
respectively.
Intended use and replay boundary
Use this dataset to audit manuscript claims, inspect frozen analyses and failure histories, reproduce provider-free validation with the linked source, or verify checksum bindings. The publication copy excludes credentials and ephemeral provider endpoints. Frozen studies must not be retried, resumed, backfilled, or rewritten to obtain a different outcome.
Do not treat these bounded tasks as evidence of general-purpose program synthesis, formal verification, a Bayesian posterior over unbounded programs, wall-clock superiority, or calibration beyond the explicitly gated Fresh V3 terminal exact-mass endpoint.
Licensing
This dataset card, experiment metadata, and sanitized run outputs are released under CC BY 4.0. Third-party model and paper references retain their original licenses. Source code remains under its repository license and is distributed through the linked GitHub revision, not this dataset.
- Downloads last month
- 782