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
record_count: int64
records_root: string
repetition_metric: struct<input: string, local: string, global: string, interpretation: string>
child 0, input: string
child 1, local: string
child 2, global: string
child 3, interpretation: string
conditions: list<item: struct<condition: string, n: int64, correct: int64, incorrect: int64, unresolved: int64, (... 459 chars omitted)
child 0, item: struct<condition: string, n: int64, correct: int64, incorrect: int64, unresolved: int64, natural_sto (... 447 chars omitted)
child 0, condition: string
child 1, n: int64
child 2, correct: int64
child 3, incorrect: int64
child 4, unresolved: int64
child 5, natural_stop_finish_reason_stop: int64
child 6, token_cap_finish_reason_length: int64
child 7, wall_time_finish_reason: int64
child 8, mean_tokens: double
child 9, median_tokens: double
child 10, max_tokens: int64
child 11, mean_max_trailing512_duplicate8_fraction: double
child 12, median_max_trailing512_duplicate8_fraction: double
child 13, n_max_window_duplicate8_fraction_ge_0_5: int64
child 14, mean_full_response_duplicate8_fraction: double
child 15, mean_max_full_response_8gram_frequency: double
child 16, eos_id_in_output_metadata_true: int64
per_question: list<item: struct<source_row: int64, base_seed0: string, base_seed1: string, base_correct_of_2: int6 (... 540 chars omitted)
child 0, item: struct<source_row: int64, base_s
...
hild 5, question_bootstrap_percentile_95: list<item: double>
child 0, item: double
child 6, first_only: int64
child 7, second_only: int64
child 8, exact_mcnemar_two_sided_p: null
child 9, per_question: list<item: struct<question_id: int64, paired_replicates: int64, first_accuracy: double, second_accur (... 33 chars omitted)
child 0, item: struct<question_id: int64, paired_replicates: int64, first_accuracy: double, second_accuracy: double (... 21 chars omitted)
child 0, question_id: int64
child 1, paired_replicates: int64
child 2, first_accuracy: double
child 3, second_accuracy: double
child 4, difference: double
child 10, bootstrap_seed: int64
child 11, bootstrap_samples: int64
child 12, interpretation_note: string
child 13, first_condition: string
child 14, comparator_condition: string
child 15, first_candidate_id: string
child 16, comparator_candidate_id: string
child 17, expected_generation_pairs: int64
child 18, complete_matched_generation_pairs: int64
child 19, incomplete_pairs: list<item: null>
child 0, item: null
child 20, analysis_uses_only_complete_pairs: bool
child 21, endpoint: string
to
{'run_id': Value('string'), 'calibration_sha256': Value('string'), 'calibration_artifact_sha256': Value('string'), 'calibration_artifact_integrity': {'self_hash_verified': Value('bool'), 'manifest_hash_present': Value('bool'), 'manifest_hash_verified': Value('bool'), 'parent_artifact_verified': Value('bool')}, 'calibration_manifest_sha256': Value('string'), 'split_manifest_sha256': Value('string'), 'heldout_rows': List(Value('int64')), 'heldout_seeds': List(Value('int64')), 'heldout_max_new_tokens': Value('int64'), 'heldout_max_generation_seconds': Value('int64'), 'heldout_gate_manifest_sha256': Value('string'), 'heldout_record_settings': {'verified_records': Value('int64'), 'expected_values': {'presence_penalty': Value('float64'), 'temperature': Value('float64'), 'top_k': Value('int64'), 'top_p': Value('float64'), 'max_new_tokens': Value('int64'), 'max_generation_seconds': Value('int64'), 'thinking': Value('bool')}, 'heldout_cap_source': Value('string')}, 'selection': {'phase': Value('string'), 'max_new_tokens': Value('int64'), 'expected_question_seeds': List(List(Value('int64'))), 'families': {'independent_constant': {'status': Value('string'), 'selected_candidate_id': Value('string'), 'blockers': List(Value('null')), 'candidates': {'independent_constant__joint_signed_mean': {'n_expected': Value('int64'), 'n_recorded': Value('int64'), 'n_missing': Value('int64'), 'missing_question_seeds': List(Value('null')), 'n_technical_errors': Value('int64'), 'technical_error_question_s
...
ion_pairs': Value('int64'), 'complete_matched_generation_pairs': Value('int64'), 'incomplete_pairs': List(Value('null')), 'analysis_uses_only_complete_pairs': Value('bool'), 'endpoint': Value('string')}}, 'edited_minus_rank1_activation_dependent': {'first_condition': Value('string'), 'comparator_condition': Value('string'), 'summary': {'n_questions': Value('int64'), 'n_paired_generations': Value('int64'), 'excluded_questions': List(Value('null')), 'unmatched_replicate_count': Value('int64'), 'mean_question_weighted_difference': Value('float64'), 'question_bootstrap_percentile_95': List(Value('float64')), 'first_only': Value('int64'), 'second_only': Value('int64'), 'exact_mcnemar_two_sided_p': Value('null'), 'per_question': List({'question_id': Value('int64'), 'paired_replicates': Value('int64'), 'first_accuracy': Value('float64'), 'second_accuracy': Value('float64'), 'difference': Value('float64')}), 'bootstrap_seed': Value('int64'), 'bootstrap_samples': Value('int64'), 'interpretation_note': Value('string'), 'first_condition': Value('string'), 'comparator_condition': Value('string'), 'first_candidate_id': Value('string'), 'comparator_candidate_id': Value('string'), 'expected_generation_pairs': Value('int64'), 'complete_matched_generation_pairs': Value('int64'), 'incomplete_pairs': List(Value('null')), 'analysis_uses_only_complete_pairs': Value('bool'), 'endpoint': Value('string')}}}}, 'heldout_used_for_selection': Value('bool'), 'direction_discovery_caveat': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
record_count: int64
records_root: string
repetition_metric: struct<input: string, local: string, global: string, interpretation: string>
child 0, input: string
child 1, local: string
child 2, global: string
child 3, interpretation: string
conditions: list<item: struct<condition: string, n: int64, correct: int64, incorrect: int64, unresolved: int64, (... 459 chars omitted)
child 0, item: struct<condition: string, n: int64, correct: int64, incorrect: int64, unresolved: int64, natural_sto (... 447 chars omitted)
child 0, condition: string
child 1, n: int64
child 2, correct: int64
child 3, incorrect: int64
child 4, unresolved: int64
child 5, natural_stop_finish_reason_stop: int64
child 6, token_cap_finish_reason_length: int64
child 7, wall_time_finish_reason: int64
child 8, mean_tokens: double
child 9, median_tokens: double
child 10, max_tokens: int64
child 11, mean_max_trailing512_duplicate8_fraction: double
child 12, median_max_trailing512_duplicate8_fraction: double
child 13, n_max_window_duplicate8_fraction_ge_0_5: int64
child 14, mean_full_response_duplicate8_fraction: double
child 15, mean_max_full_response_8gram_frequency: double
child 16, eos_id_in_output_metadata_true: int64
per_question: list<item: struct<source_row: int64, base_seed0: string, base_seed1: string, base_correct_of_2: int6 (... 540 chars omitted)
child 0, item: struct<source_row: int64, base_s
...
hild 5, question_bootstrap_percentile_95: list<item: double>
child 0, item: double
child 6, first_only: int64
child 7, second_only: int64
child 8, exact_mcnemar_two_sided_p: null
child 9, per_question: list<item: struct<question_id: int64, paired_replicates: int64, first_accuracy: double, second_accur (... 33 chars omitted)
child 0, item: struct<question_id: int64, paired_replicates: int64, first_accuracy: double, second_accuracy: double (... 21 chars omitted)
child 0, question_id: int64
child 1, paired_replicates: int64
child 2, first_accuracy: double
child 3, second_accuracy: double
child 4, difference: double
child 10, bootstrap_seed: int64
child 11, bootstrap_samples: int64
child 12, interpretation_note: string
child 13, first_condition: string
child 14, comparator_condition: string
child 15, first_candidate_id: string
child 16, comparator_candidate_id: string
child 17, expected_generation_pairs: int64
child 18, complete_matched_generation_pairs: int64
child 19, incomplete_pairs: list<item: null>
child 0, item: null
child 20, analysis_uses_only_complete_pairs: bool
child 21, endpoint: string
to
{'run_id': Value('string'), 'calibration_sha256': Value('string'), 'calibration_artifact_sha256': Value('string'), 'calibration_artifact_integrity': {'self_hash_verified': Value('bool'), 'manifest_hash_present': Value('bool'), 'manifest_hash_verified': Value('bool'), 'parent_artifact_verified': Value('bool')}, 'calibration_manifest_sha256': Value('string'), 'split_manifest_sha256': Value('string'), 'heldout_rows': List(Value('int64')), 'heldout_seeds': List(Value('int64')), 'heldout_max_new_tokens': Value('int64'), 'heldout_max_generation_seconds': Value('int64'), 'heldout_gate_manifest_sha256': Value('string'), 'heldout_record_settings': {'verified_records': Value('int64'), 'expected_values': {'presence_penalty': Value('float64'), 'temperature': Value('float64'), 'top_k': Value('int64'), 'top_p': Value('float64'), 'max_new_tokens': Value('int64'), 'max_generation_seconds': Value('int64'), 'thinking': Value('bool')}, 'heldout_cap_source': Value('string')}, 'selection': {'phase': Value('string'), 'max_new_tokens': Value('int64'), 'expected_question_seeds': List(List(Value('int64'))), 'families': {'independent_constant': {'status': Value('string'), 'selected_candidate_id': Value('string'), 'blockers': List(Value('null')), 'candidates': {'independent_constant__joint_signed_mean': {'n_expected': Value('int64'), 'n_recorded': Value('int64'), 'n_missing': Value('int64'), 'missing_question_seeds': List(Value('null')), 'n_technical_errors': Value('int64'), 'technical_error_question_s
...
ion_pairs': Value('int64'), 'complete_matched_generation_pairs': Value('int64'), 'incomplete_pairs': List(Value('null')), 'analysis_uses_only_complete_pairs': Value('bool'), 'endpoint': Value('string')}}, 'edited_minus_rank1_activation_dependent': {'first_condition': Value('string'), 'comparator_condition': Value('string'), 'summary': {'n_questions': Value('int64'), 'n_paired_generations': Value('int64'), 'excluded_questions': List(Value('null')), 'unmatched_replicate_count': Value('int64'), 'mean_question_weighted_difference': Value('float64'), 'question_bootstrap_percentile_95': List(Value('float64')), 'first_only': Value('int64'), 'second_only': Value('int64'), 'exact_mcnemar_two_sided_p': Value('null'), 'per_question': List({'question_id': Value('int64'), 'paired_replicates': Value('int64'), 'first_accuracy': Value('float64'), 'second_accuracy': Value('float64'), 'difference': Value('float64')}), 'bootstrap_seed': Value('int64'), 'bootstrap_samples': Value('int64'), 'interpretation_note': Value('string'), 'first_condition': Value('string'), 'comparator_condition': Value('string'), 'first_candidate_id': Value('string'), 'comparator_candidate_id': Value('string'), 'expected_generation_pairs': Value('int64'), 'complete_matched_generation_pairs': Value('int64'), 'incomplete_pairs': List(Value('null')), 'analysis_uses_only_complete_pairs': Value('bool'), 'endpoint': Value('string')}}}}, 'heldout_used_for_selection': Value('bool'), 'direction_discovery_caveat': 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.
ModelClinic Experiment 001: Complete Available Research Archive
This dataset contains the complete available ModelClinic Experiment 001 research archive, with documented exclusions, spanning Experiments 001A–001D. It preserves configurations, preregistrations, retained raw records, analysis code and outputs, execution failures and recoveries, resource records, reports, and reproducible figures. The earlier A/B-only public release remains intact as its own historical revision; this version adds the 001C/001D record and the final audit.
What the evidence says
- 001A, structure: the published edit changes two MLP output projections and 130,188 BF16 values. Each stored difference is overwhelmingly concentrated in its leading singular component, though not exactly rank one. The full checkpoint weights are not mirrored.
- 001B, behavior: on ten MATH-500 questions sampled twice per condition, the edited checkpoint scored 8/20 and a calibrated published-direction constant scored 7/20. The equal-question difference was +0.05, with a 95% question-cluster interval of [0.00, 0.15]. This small screen establishes neither superiority nor equivalence.
- 001C, teacher-forced likelihood: on 455 MATH-500 questions, donor-shuffled minus dynamic gate NLL was −0.000274985 nats/token, 95% interval [−0.001747665, +0.001207439]. The 90% interval met the preregistered ±0.01 margin on this narrow outcome. The margin lacks a behavioral justification; no base-NLL arm, known-effect positive control, or free-generation endpoint was included.
- 001D, fixed-prefix behavior: authentic dynamic gating produced 5/20 correct natural completions versus 2/20 for donor gating on ten selected repetitive GSM8K prefixes. The dynamic-minus-donor difference was +0.15, with 95% question-cluster interval [−0.10, +0.40]. The donor arm's realized perturbations were substantially larger after branching. This does not isolate context alignment or establish a causal advantage.
The strongest supported interpretation is that the edit has an activation-dependent numerical gate, while the behavioral importance of that gate remains unresolved. Experiment 002 should use a validated online control that matches current-step perturbation magnitude while changing gate sign/alignment.
Read the research
- Definitive technical article
- Technical appendix
- Claim-by-claim scientific audit
- Publication and availability report
- Analysis-only reproduction commands
- Data provenance and availability categories
- Limitations
Contents
experiments/: split definitions, frozen manifests, preregistrations, and pinned GSM8K test data. The 001C MATH-500 source dump is intentionally withheld because the pinned HF subset card has no license field.results/: retained 001A/001B outputs plus the 001C teacher-forcing archive and full retained 001D attempts, raw records, recoveries, analysis, tokenizer files, and resource accounting.code/: analysis, scoring, plotting, and runtime implementations for the released stages.reports/: original and corrected scientific reports, controller and validity audits, and final release record.figures/experiment_001b/andfigures/experiment_001_final/: figures, generated source CSVs, and plotting scripts.logs/: prior execution records, sanitized for public release.
Reproduction and costs
Analysis-only commands use released records and require no API credentials or GPU. The 001D analysis decodes retained tokens using a small Qwen tokenizer/configuration subset included under Apache-2.0. Figure regeneration uses CPU. Fresh model inference is a separate optional reproduction requiring pinned model checkpoints, compatible hardware, and external compute; it can incur charges and was not run for this release.
Licensing and provenance
The dataset card uses license: other because components have different terms. See LICENSE.md, licenses/, and DATA_PROVENANCE.md. The repository does not mirror model weights or original authors' full logs. ModelClinic is independent of the original publication and is not affiliated with or endorsed by MIT CSAIL or the original researchers.
- Downloads last month
- 317