Datasets:

arcadia-impact
/

python4-gemma3-27b-generalization

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
condition: string
episode: struct<benchmark_cell: string, difficulty: string, gold_python4: string, gold_rule_tags: struct<end_ (... 1885 chars omitted)
  child 0, benchmark_cell: string
  child 1, difficulty: string
  child 2, gold_python4: string
  child 3, gold_rule_tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, manual_allocation: bool (... 151 chars omitted)
      child 0, end_inclusive_slice: bool
      child 1, grouped_large_integer: bool
      child 2, lambda: bool
      child 3, manual_allocation: bool
      child 4, negative_exclusion: bool
      child 5, one_based_positive_indexing: bool
      child 6, out_parameter: bool
      child 7, statement_terminators: bool
      child 8, uppercase_boolean: bool
      child 9, walrus: bool
  child 4, held_out_rules: list<item: string>
      child 0, item: string
  child 5, parameter_names: list<item: string>
      child 0, item: string
  child 6, problem: string
  child 7, problem_id: string
  child 8, reference_rule_tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, negative_exclusion: boo (... 76 chars omitted)
      child 0, end_inclusive_slice: bool
      child 1, grouped_large_integer: bool
      child 2, lambda: bool
      child 3, negative_exclusion: bool
      child 4, one_based_positive_indexing: bool
      child 5, uppercase_boolean: bool
      child 6, walrus: bool
  child 9, source_row_sha256: string
  child 10, source_split: string
  child 11, te
...
        child 61, number: string
              child 62, fill: string
              child 63, numbers: list<item: string>
                  child 0, item: string
format_valid: bool
probe: string
python3: struct<error_kind: string, python3_compile: bool, python3_pass: bool, stderr: string, stdout: string (... 1 chars omitted)
  child 0, error_kind: string
  child 1, python3_compile: bool
  child 2, python3_pass: bool
  child 3, stderr: string
  child 4, stdout: string
python4: struct<boa_compile: bool, boa_pass: bool, error_kind: string, python4_adoption: bool, rule_pass: str (... 317 chars omitted)
  child 0, boa_compile: bool
  child 1, boa_pass: bool
  child 2, error_kind: string
  child 3, python4_adoption: bool
  child 4, rule_pass: struct<>
  child 5, stderr: string
  child 6, stdout: string
  child 7, tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, manual_allocation: bool (... 151 chars omitted)
      child 0, end_inclusive_slice: bool
      child 1, grouped_large_integer: bool
      child 2, lambda: bool
      child 3, manual_allocation: bool
      child 4, negative_exclusion: bool
      child 5, one_based_positive_indexing: bool
      child 6, out_parameter: bool
      child 7, statement_terminators: bool
      child 8, uppercase_boolean: bool
      child 9, walrus: bool
  child 8, warning_free: bool
response: string
system: string
task_id: string
arm: string
completed_at: string
conditions: list<item: string>
  child 0, item: string
to
{'completed_at': Value('string'), 'arm': Value('string'), 'conditions': List(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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              condition: string
              episode: struct<benchmark_cell: string, difficulty: string, gold_python4: string, gold_rule_tags: struct<end_ (... 1885 chars omitted)
                child 0, benchmark_cell: string
                child 1, difficulty: string
                child 2, gold_python4: string
                child 3, gold_rule_tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, manual_allocation: bool (... 151 chars omitted)
                    child 0, end_inclusive_slice: bool
                    child 1, grouped_large_integer: bool
                    child 2, lambda: bool
                    child 3, manual_allocation: bool
                    child 4, negative_exclusion: bool
                    child 5, one_based_positive_indexing: bool
                    child 6, out_parameter: bool
                    child 7, statement_terminators: bool
                    child 8, uppercase_boolean: bool
                    child 9, walrus: bool
                child 4, held_out_rules: list<item: string>
                    child 0, item: string
                child 5, parameter_names: list<item: string>
                    child 0, item: string
                child 6, problem: string
                child 7, problem_id: string
                child 8, reference_rule_tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, negative_exclusion: boo (... 76 chars omitted)
                    child 0, end_inclusive_slice: bool
                    child 1, grouped_large_integer: bool
                    child 2, lambda: bool
                    child 3, negative_exclusion: bool
                    child 4, one_based_positive_indexing: bool
                    child 5, uppercase_boolean: bool
                    child 6, walrus: bool
                child 9, source_row_sha256: string
                child 10, source_split: string
                child 11, te
              ...
                      child 61, number: string
                            child 62, fill: string
                            child 63, numbers: list<item: string>
                                child 0, item: string
              format_valid: bool
              probe: string
              python3: struct<error_kind: string, python3_compile: bool, python3_pass: bool, stderr: string, stdout: string (... 1 chars omitted)
                child 0, error_kind: string
                child 1, python3_compile: bool
                child 2, python3_pass: bool
                child 3, stderr: string
                child 4, stdout: string
              python4: struct<boa_compile: bool, boa_pass: bool, error_kind: string, python4_adoption: bool, rule_pass: str (... 317 chars omitted)
                child 0, boa_compile: bool
                child 1, boa_pass: bool
                child 2, error_kind: string
                child 3, python4_adoption: bool
                child 4, rule_pass: struct<>
                child 5, stderr: string
                child 6, stdout: string
                child 7, tags: struct<end_inclusive_slice: bool, grouped_large_integer: bool, lambda: bool, manual_allocation: bool (... 151 chars omitted)
                    child 0, end_inclusive_slice: bool
                    child 1, grouped_large_integer: bool
                    child 2, lambda: bool
                    child 3, manual_allocation: bool
                    child 4, negative_exclusion: bool
                    child 5, one_based_positive_indexing: bool
                    child 6, out_parameter: bool
                    child 7, statement_terminators: bool
                    child 8, uppercase_boolean: bool
                    child 9, walrus: bool
                child 8, warning_free: bool
              response: string
              system: string
              task_id: string
              arm: string
              completed_at: string
              conditions: list<item: string>
                child 0, item: string
              to
              {'completed_at': Value('string'), 'arm': Value('string'), 'conditions': List(Value('string'))}
              because column names don't match

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.

Gemma 3 27B Python4 defaultization evaluation

This dataset contains raw prompts, prompt audits, model responses, mechanical grades, summaries, configs, source manifests, and logs for final run 20260812T235200Z-generalization-v1.

The evaluation compares five midtraining parents under four conditions:

Floor: bare parent, ordinary Python request.
AFT: rank-64 90:10 Python4:Dolci adapter, same ordinary Python request.
RL: rank-64 AFT+RL continuation, same ordinary Python request.
Name cue: bare parent with only the language name changed to Python4 and no rule description. This is a cue diagnostic, not a performance upper bound.

Prompt-audit records confirm that Floor, AFT, and RL are byte-identical and contain no model-visible mention of Python4, Python3, Boa, or a dialect rule.

Natural-task success

Model	Floor	AFT	AFT+RL	Name cue	AFT held-out	AFT+RL held-out
Control	0/128	35/128	36/128	0/128	21/96	23/96
1ep Midtrain	0/128	36/128	26/128	0/128	25/96	15/96
1ep SDF	0/128	28/128	26/128	0/128	20/96	14/96
4ep Midtrain	0/128	32/128	37/128	0/128	19/96	24/96
4ep SDF	0/128	36/128	36/128	0/128	23/96	24/96

Corrected indexing diagnostics

Every slice probe has a positive lower and upper bound and four distinct answers: one-based/inclusive (Python4), zero-based/exclusive (Python3), one-based/exclusive, and zero-based/inclusive. No xs[:3] or xs[::-1] probe can receive Python4 credit. Exclusion requires xs[-k] to remove the kth one-based element; Python3 scalar from-end lookup and zero-based removal are separate outcomes. All 16 gold answers pass Boa. Formatting does not gate semantic correctness.

Model	Ordinary parent slice/exclusion	AFT	AFT+RL	Name-cued parent
Control	0/8, 0/8	0/8, 0/8	0/8, 0/8	0/8, 0/8
1ep Midtrain	0/8, 0/8	0/8, 0/8	0/8, 0/8	0/8, 0/8
1ep SDF	1/8, 0/8	3/8, 0/8	0/8, 0/8	1/8, 0/8
4ep Midtrain	0/8, 0/8	3/8, 2/8	1/8, 0/8	4/8, 0/8
4ep SDF	5/8, 0/8	4/8, 2/8	6/8, 2/8	6/8, 1/8

The corrected Control is 0/8 throughout. Four-epoch SDF shows the clearest uncued slice generalization. Exact exclusion remains rare.

Provenance

Evaluation launch commit: 92bc03214c3c0d394ae6b3348ec2656a7e01d527.
Parent repository revision: arcadia-impact/python4-gemma3-27b@415ce4d73de6ed42b1cb3ee196909655dda8138d.
AFT revision: arcadia-impact/python4-gemma3-27b-aft@79c3ed038ae06267c745e6d49d2a988f76ee5436.
AFT+RL adapter revision: arcadia-impact/python4-gemma3-27b-rlvr@e90f985fe34b6a75e5f2252899b5a4da7b49e88c.
Boa revision: a215d2d1875f3d3d986185597c7f12a1d0258568.

The consolidated results.csv includes this run alongside the earlier AFT, RLVR, Q/A, and standard evaluations. The final defaultization and updated rule figures are included as PDFs.

Downloads last month: -

Total file size:

18.7 MB