Datasets:

queelius
/

rlm-oolong-split-and-verdicts

Tags:

License:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Preview

Duplicate

Split (1)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed

Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
evaluation: string
controller: string
policy: string
id: int64
context_len: int64
task: string
correct: bool
parsed: string
total_tokens: int64
prompt_tokens: int64
completion_tokens: int64
calls: int64
turns: int64
stop_reason: string
error: string
context_window_id: int64
dataset: string
split: string
task_group: string
answer_type: string
to
{'id': Value('int64'), 'context_len': Value('int64'), 'context_window_id': Value('int64'), 'dataset': Value('string'), 'task_group': Value('string'), 'task': Value('string'), 'answer_type': Value('string'), 'split': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                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
              evaluation: string
              controller: string
              policy: string
              id: int64
              context_len: int64
              task: string
              correct: bool
              parsed: string
              total_tokens: int64
              prompt_tokens: int64
              completion_tokens: int64
              calls: int64
              turns: int64
              stop_reason: string
              error: string
              context_window_id: int64
              dataset: string
              split: string
              task_group: string
              answer_type: string
              to
              {'id': Value('int64'), 'context_len': Value('int64'), 'context_window_id': Value('int64'), 'dataset': Value('string'), 'task_group': Value('string'), 'task': Value('string'), 'answer_type': Value('string'), 'split': Value('string')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

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.

id int64	context_len int64	context_window_id int64	dataset string	task_group string	task string	answer_type string	split string
110,010,000	1,024	10,000	spam	counting	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,001	1,024	10,000	spam	counting	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,002	1,024	10,000	spam	counting	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
110,010,003	1,024	10,000	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,004	1,024	10,000	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,005	1,024	10,000	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
110,010,006	1,024	10,000	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,007	1,024	10,000	spam	user	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,008	1,024	10,000	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
110,010,009	1,024	10,000	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,010	1,024	10,000	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,011	1,024	10,000	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
110,010,012	1,024	10,000	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	teach
110,010,013	1,024	10,000	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
110,010,014	1,024	10,000	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	teach
110,010,015	1,024	10,000	spam	timeline	TASK_TYPE.REPRESENTED_N_TIMES	ANSWER_TYPE.NUMERIC	teach
110,010,016	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
110,010,017	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
110,010,018	1,024	10,000	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,019	1,024	10,000	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,020	1,024	10,000	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,021	1,024	10,000	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,022	1,024	10,000	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
110,010,023	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
110,010,024	1,024	10,000	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,025	1,024	10,000	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
110,010,026	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.MONTH_YEAR	teach
110,010,027	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.NUMERIC	teach
110,010,028	1,024	10,000	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.NUMERIC	teach
110,010,029	1,024	10,000	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.NUMERIC	teach
110,010,030	1,024	10,000	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.NUMERIC	teach
110,010,031	1,024	10,001	spam	counting	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,032	1,024	10,001	spam	counting	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,033	1,024	10,001	spam	counting	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
110,010,034	1,024	10,001	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
110,010,035	1,024	10,001	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
110,010,036	1,024	10,001	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	held_out
110,010,037	1,024	10,001	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,038	1,024	10,001	spam	user	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,039	1,024	10,001	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
110,010,040	1,024	10,001	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	held_out
110,010,041	1,024	10,001	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	held_out
110,010,042	1,024	10,001	spam	timeline	TASK_TYPE.REPRESENTED_N_TIMES	ANSWER_TYPE.NUMERIC	held_out
110,010,043	1,024	10,001	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
110,010,044	1,024	10,001	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,045	1,024	10,001	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
110,010,046	1,024	10,001	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
110,010,047	1,024	10,001	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.MONTH_YEAR	held_out
110,010,048	1,024	10,001	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.NUMERIC	held_out
110,010,049	1,024	10,001	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.NUMERIC	held_out
111,010,003	2,048	10,003	spam	counting	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,004	2,048	10,003	spam	counting	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,005	2,048	10,003	spam	counting	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
111,010,006	2,048	10,003	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
111,010,007	2,048	10,003	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
111,010,008	2,048	10,003	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	held_out
111,010,009	2,048	10,003	spam	user	TASK_TYPE.SECOND_MOST_FREQ	ANSWER_TYPE.USER	held_out
111,010,010	2,048	10,003	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,011	2,048	10,003	spam	user	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,012	2,048	10,003	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
111,010,013	2,048	10,003	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	held_out
111,010,014	2,048	10,003	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	held_out
111,010,015	2,048	10,003	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	held_out
111,010,016	2,048	10,003	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	held_out
111,010,017	2,048	10,003	spam	timeline	TASK_TYPE.REPRESENTED_N_TIMES	ANSWER_TYPE.NUMERIC	held_out
111,010,018	2,048	10,003	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
111,010,019	2,048	10,003	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
111,010,020	2,048	10,003	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,021	2,048	10,003	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,022	2,048	10,003	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
111,010,023	2,048	10,003	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
111,010,024	2,048	10,003	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	held_out
111,010,025	2,048	10,003	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,026	2,048	10,003	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	held_out
111,010,027	2,048	10,003	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	held_out
111,010,028	2,048	10,004	spam	counting	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,029	2,048	10,004	spam	counting	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,030	2,048	10,004	spam	counting	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
111,010,031	2,048	10,004	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,032	2,048	10,004	spam	counting	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,033	2,048	10,004	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
111,010,034	2,048	10,004	spam	user	TASK_TYPE.SECOND_MOST_FREQ	ANSWER_TYPE.USER	teach
111,010,035	2,048	10,004	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,036	2,048	10,004	spam	user	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,037	2,048	10,004	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
111,010,038	2,048	10,004	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,039	2,048	10,004	spam	user	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,040	2,048	10,004	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
111,010,041	2,048	10,004	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	teach
111,010,042	2,048	10,004	spam	user	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.USER	teach
111,010,043	2,048	10,004	spam	user	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.USER	teach
111,010,044	2,048	10,004	spam	timeline	TASK_TYPE.REPRESENTED_N_TIMES	ANSWER_TYPE.NUMERIC	teach
111,010,045	2,048	10,004	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
111,010,046	2,048	10,004	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
111,010,047	2,048	10,004	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,048	2,048	10,004	spam	timeline	TASK_TYPE.LEAST_FREQ	ANSWER_TYPE.LABEL	teach
111,010,049	2,048	10,004	spam	timeline	TASK_TYPE.RELATIVE_FREQ	ANSWER_TYPE.COMPARISON	teach
111,010,050	2,048	10,004	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,051	2,048	10,004	spam	timeline	TASK_TYPE.NUMERIC_ONE_CLASS	ANSWER_TYPE.NUMERIC	teach
111,010,052	2,048	10,004	spam	timeline	TASK_TYPE.MOST_FREQ	ANSWER_TYPE.LABEL	teach

End of preview.

rlm-oolong-split-and-verdicts

Held-out split definition and per-row verdicts for RLM controllers on OOLONG-synth (<=32k tokens).

split.jsonl (600 rows): for every OOLONG-synth validation row up to 32k tokens, its id, context_len, context_window_id and our split label (teach = even window ids used for teaching and RL, held_out = odd; alternating ids at 4096 tokens where only one parity exists). verdicts.jsonl (1758 rows): one row per evaluated answer for each full held-out evaluation (evaluation names the run: direct answering, SFT-005, and the four RL controllers), with correct, the parsed answer, endpoint-reported tokens, helper calls, turns and errors. Verifier: format-tolerant exact match on the answer line (oolong_eval.verify). These rows reproduce every OOLONG number in docs/RESULTS.md of https://github.com/queelius/rlm. No benchmark text is included (see the source dataset oolongbench/oolong-synth, revision f0d59eaf, which declares no license).

Citation: Towell, A. (2026). RLM research repository, https://github.com/queelius/rlm.

Downloads last month: 14

Total file size:

715 kB

Models trained or fine-tuned on queelius/rlm-oolong-split-and-verdicts

queelius/rlm-qwen3-4b-sft-005

Updated 2 days ago • 13

queelius/rlm-qwen3-4b-rl-mixed-001

Updated 2 days ago • 12