Dataset Viewer
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: CastError
Message: Couldn't cast
schema: string
study_date: timestamp[s]
started_at: string
completed_at: string
scope: string
provenance: struct<binary_sha256: string, bridge_sha256: string, binary_and_bridge_unchanged: bool, inputs_uncha (... 11 chars omitted)
child 0, binary_sha256: string
child 1, bridge_sha256: string
child 2, binary_and_bridge_unchanged: bool
child 3, inputs_unchanged: bool
summary: struct<total_evaluations: int64, failed_evaluations: int64, wall_seconds: double, remote_model_calls (... 35 chars omitted)
child 0, total_evaluations: int64
child 1, failed_evaluations: int64
child 2, wall_seconds: double
child 3, remote_model_calls: int64
child 4, mechanism_qualified: bool
aggregates: list<item: struct<variant: string, selected_inputs: int64, successful_evaluations: int64, failed_eva (... 927 chars omitted)
child 0, item: struct<variant: string, selected_inputs: int64, successful_evaluations: int64, failed_evaluations: i (... 915 chars omitted)
child 0, variant: string
child 1, selected_inputs: int64
child 2, successful_evaluations: int64
child 3, failed_evaluations: int64
child 4, plans: int64
child 5, no_plan: int64
child 6, sum_context_before: int64
child 7, sum_projected_reclaimed: int64
child 8, median_projected_reduction_percent_including_noops: double
child 9, min_projected_reduction_percent: double
child 10, max_projected_reduction_percent: double
child 11, literal_samples: int64
chil
...
5, sum_context_before: int64
child 16, sum_projected_reclaimed: int64
child 17, baseline_unavailable_samples: int64
child 18, baseline_error_samples: int64
child 19, planned_verification_delta_available: int64
child 20, planned_verification_delta_unavailable: int64
child 21, post_error_samples: int64
child 22, new_error_count: int64
child 23, new_warning_count: int64
child 24, nonzero_probe_samples: int64
child 25, zero_probe_samples: int64
child 26, probe_score_unavailable_plans: int64
child 27, literal_probes_total: int64
child 28, literal_probes_retained: int64
child 29, nonzero_tail_probe_samples: int64
child 30, tail_probes_total: int64
child 31, tail_probes_retained: int64
child 32, median_wall_ms: double
child 6, high_context_sessions: int64
child 7, below_trigger_controls: int64
binary_sha256: string
combined: struct<selected_sessions: int64, strategy_evaluations: int64, evaluation_failures: int64, compacted_ (... 100 chars omitted)
child 0, selected_sessions: int64
child 1, strategy_evaluations: int64
child 2, evaluation_failures: int64
child 3, compacted_plans: int64
child 4, compacted_no_plan: int64
child 5, compacted_median_reduction_including_no_plan_percent: int64
measurement: string
source_commit: string
to
{'schema': Value('string'), 'study_date': Value('timestamp[s]'), 'measurement': Value('string'), 'source_commit': Value('string'), 'binary_sha256': Value('string'), 'combined': {'selected_sessions': Value('int64'), 'strategy_evaluations': Value('int64'), 'evaluation_failures': Value('int64'), 'compacted_plans': Value('int64'), 'compacted_no_plan': Value('int64'), 'compacted_median_reduction_including_no_plan_percent': Value('int64')}, 'cohorts': List({'name': Value('string'), 'selected_sessions': Value('int64'), 'codex_subagent_sessions': Value('int64'), 'claude_sessions': Value('int64'), 'wall_seconds': Value('float64'), 'aggregates': List({'provider': Value('string'), 'cohort': Value('string'), 'strategy': Value('string'), 'stratum': Value('string'), 'selected_samples': Value('int64'), 'successful_evals': Value('int64'), 'plans': Value('int64'), 'no_plan': Value('int64'), 'failures': Value('int64'), 'primary_reduction_samples': Value('int64'), 'median_reduction_including_no_plan_percent': Value('float64'), 'median_reduction_planned_only_percent': Value('float64'), 'min_reduction_including_no_plan_percent': Value('float64'), 'max_reduction_including_no_plan_percent': Value('float64'), 'median_descriptive_resampling_95pct_interval': List(Value('float64')), 'sum_context_before': Value('int64'), 'sum_projected_reclaimed': Value('int64'), 'baseline_unavailable_samples': Value('int64'), 'baseline_error_samples': Value('int64'), 'planned_verification_delta_available': Value('int64'), 'planned_verification_delta_unavailable': Value('int64'), 'post_error_samples': Value('int64'), 'new_error_count': Value('int64'), 'new_warning_count': Value('int64'), 'nonzero_probe_samples': Value('int64'), 'zero_probe_samples': Value('int64'), 'probe_score_unavailable_plans': Value('int64'), 'literal_probes_total': Value('int64'), 'literal_probes_retained': Value('int64'), 'nonzero_tail_probe_samples': Value('int64'), 'tail_probes_total': Value('int64'), 'tail_probes_retained': Value('int64'), 'median_wall_ms': Value('float64')}), 'high_context_sessions': Value('int64'), 'below_trigger_controls': Value('int64')}), 'privacy': Value('string'), 'limitations': List(Value('string'))}
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
schema: string
study_date: timestamp[s]
started_at: string
completed_at: string
scope: string
provenance: struct<binary_sha256: string, bridge_sha256: string, binary_and_bridge_unchanged: bool, inputs_uncha (... 11 chars omitted)
child 0, binary_sha256: string
child 1, bridge_sha256: string
child 2, binary_and_bridge_unchanged: bool
child 3, inputs_unchanged: bool
summary: struct<total_evaluations: int64, failed_evaluations: int64, wall_seconds: double, remote_model_calls (... 35 chars omitted)
child 0, total_evaluations: int64
child 1, failed_evaluations: int64
child 2, wall_seconds: double
child 3, remote_model_calls: int64
child 4, mechanism_qualified: bool
aggregates: list<item: struct<variant: string, selected_inputs: int64, successful_evaluations: int64, failed_eva (... 927 chars omitted)
child 0, item: struct<variant: string, selected_inputs: int64, successful_evaluations: int64, failed_evaluations: i (... 915 chars omitted)
child 0, variant: string
child 1, selected_inputs: int64
child 2, successful_evaluations: int64
child 3, failed_evaluations: int64
child 4, plans: int64
child 5, no_plan: int64
child 6, sum_context_before: int64
child 7, sum_projected_reclaimed: int64
child 8, median_projected_reduction_percent_including_noops: double
child 9, min_projected_reduction_percent: double
child 10, max_projected_reduction_percent: double
child 11, literal_samples: int64
chil
...
5, sum_context_before: int64
child 16, sum_projected_reclaimed: int64
child 17, baseline_unavailable_samples: int64
child 18, baseline_error_samples: int64
child 19, planned_verification_delta_available: int64
child 20, planned_verification_delta_unavailable: int64
child 21, post_error_samples: int64
child 22, new_error_count: int64
child 23, new_warning_count: int64
child 24, nonzero_probe_samples: int64
child 25, zero_probe_samples: int64
child 26, probe_score_unavailable_plans: int64
child 27, literal_probes_total: int64
child 28, literal_probes_retained: int64
child 29, nonzero_tail_probe_samples: int64
child 30, tail_probes_total: int64
child 31, tail_probes_retained: int64
child 32, median_wall_ms: double
child 6, high_context_sessions: int64
child 7, below_trigger_controls: int64
binary_sha256: string
combined: struct<selected_sessions: int64, strategy_evaluations: int64, evaluation_failures: int64, compacted_ (... 100 chars omitted)
child 0, selected_sessions: int64
child 1, strategy_evaluations: int64
child 2, evaluation_failures: int64
child 3, compacted_plans: int64
child 4, compacted_no_plan: int64
child 5, compacted_median_reduction_including_no_plan_percent: int64
measurement: string
source_commit: string
to
{'schema': Value('string'), 'study_date': Value('timestamp[s]'), 'measurement': Value('string'), 'source_commit': Value('string'), 'binary_sha256': Value('string'), 'combined': {'selected_sessions': Value('int64'), 'strategy_evaluations': Value('int64'), 'evaluation_failures': Value('int64'), 'compacted_plans': Value('int64'), 'compacted_no_plan': Value('int64'), 'compacted_median_reduction_including_no_plan_percent': Value('int64')}, 'cohorts': List({'name': Value('string'), 'selected_sessions': Value('int64'), 'codex_subagent_sessions': Value('int64'), 'claude_sessions': Value('int64'), 'wall_seconds': Value('float64'), 'aggregates': List({'provider': Value('string'), 'cohort': Value('string'), 'strategy': Value('string'), 'stratum': Value('string'), 'selected_samples': Value('int64'), 'successful_evals': Value('int64'), 'plans': Value('int64'), 'no_plan': Value('int64'), 'failures': Value('int64'), 'primary_reduction_samples': Value('int64'), 'median_reduction_including_no_plan_percent': Value('float64'), 'median_reduction_planned_only_percent': Value('float64'), 'min_reduction_including_no_plan_percent': Value('float64'), 'max_reduction_including_no_plan_percent': Value('float64'), 'median_descriptive_resampling_95pct_interval': List(Value('float64')), 'sum_context_before': Value('int64'), 'sum_projected_reclaimed': Value('int64'), 'baseline_unavailable_samples': Value('int64'), 'baseline_error_samples': Value('int64'), 'planned_verification_delta_available': Value('int64'), 'planned_verification_delta_unavailable': Value('int64'), 'post_error_samples': Value('int64'), 'new_error_count': Value('int64'), 'new_warning_count': Value('int64'), 'nonzero_probe_samples': Value('int64'), 'zero_probe_samples': Value('int64'), 'probe_score_unavailable_plans': Value('int64'), 'literal_probes_total': Value('int64'), 'literal_probes_retained': Value('int64'), 'nonzero_tail_probe_samples': Value('int64'), 'tail_probes_total': Value('int64'), 'tail_probes_retained': Value('int64'), 'median_wall_ms': Value('float64')}), 'high_context_sessions': Value('int64'), 'below_trigger_controls': Value('int64')}), 'privacy': Value('string'), 'limitations': List(Value('string'))}
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.
Gobstopper benchmark reports
Public aggregate results and protocols from Gobstopper's compaction studies. Use the downloads to inspect selection rules, denominators, retention measurements, recovery tests, and the limits of each historical experiment.
Studies
data/2026-09-19/ contains the retrospective's aggregate results, protocol, and report, plus the separately recorded retention ablation and Apple retention pilot. Keep each experiment's corpus, policy, executable hashes, and limitations with its results. Projected token reduction is an offline estimate. It does not measure provider billing or successful continuation of a task.
data/2026-09-20/ contains the recovery study's protocol and results. Its synthetic fixture evaluation is separate from the private-session studies. Follow the public reproduction instructions in the protocol and the source repository.
The private studies came from one person's machine. Their published aggregates cannot establish performance across other users or independently reproduce the private corpus. Literal string retention does not establish semantic importance or task quality. Historical results describe their recorded implementations, not every later Gobstopper release or request-time proxy strategy.
These downloads form an artifact archive with several JSON schemas. Read individual files using their recorded schema; a combined datasets.load_dataset table is not provided.
Privacy and provenance
Only the explicitly selected public reports and protocols are included. Private transcript text, session identifiers, local paths, vault snapshots, per-session hashes, and per-session results are excluded. Do not add those files when extending this collection.
export-manifest.json records the source commit and SHA-256 of every exported artifact. New completed studies receive new dated folders. Published study files remain frozen; corrections receive a new version with an explanation of the change. See Gobstopper for current product behavior.
License and attribution
The selected aggregate reports and protocols in this archive are by Hraness and licensed under Creative Commons Attribution 4.0 International. Attribute Hraness and link to the original benchmark page and the source revision recorded in export-manifest.json.
This license covers the selected files listed in export-manifest.json and this card. It does not cover source code, private transcripts, or raw provider responses. Source code retains its original terms.
- Downloads last month
- 38