The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column(/stream_40k/6gram_masked_top/[]/[]) changed from string to number in row 0
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 304, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
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/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Pascaline-3M synthetic arithmetic: a public sample
5,000 problems from the synthetic arithmetic stream used in the decay phase of Pascaline-3M, where it made up 30% of the 1.2B decay tokens. The generators are our own programs: every answer is computed exactly, and nothing is written by an LLM. This sample shows what the data looks like and lets anyone check it against ArithMark-3. The generator code itself is not published.
Contents
sample.jsonl: one problem per line,{"id", "generator", "text"}. Equal parts from the four generators, which training mixed 25% each:
generator |
What it covers |
|---|---|
v2_breadth |
Broad grade-school topics: money and change, time, lengths, tallies, sequences, arrays, shapes, fractions, comparisons, bare equations |
v3_stories |
Compositional gain/loss stories with operands balanced by carry/borrow type, short-result subtraction, clock arithmetic |
v4_relations |
"N more/fewer than X", differences asked many ways, sharing/filling/grouping division, combined and compared measures, three-number problems. 20% minimal pairs (same numbers, opposite relation), 20% with an equation tail |
v4.1_multistep |
2–4-step chains with distractor sentences, two-hop relations, gaps after a change, multi-step leftovers, perimeters |
fairness_report.json: overlap with all 1,000 ArithMark-3 rows, measured on this sample and on a larger 40,000-problem draw of the same stream.
Each generator composes its phrasings from parts split into a training set and a held-out set. This sample (like training) uses only the training parts; the held-out phrasings are kept for our own evaluation. Documents in training were groups of 1–8 such problems separated by newlines.
Fairness checks
Measured against every ArithMark-3 row (its context alone, and its context plus each ending), as lower-cased word n-grams:
| Check | Sample (5k) | Stream draw (40k) |
|---|---|---|
| Exact 13-gram overlap | 0% | 0% |
| Exact 8-gram overlap | 0% | 0% |
| 6-gram overlap with numbers masked | 0% | 0.003% |
| Banned ArithMark sentence frames (original list, 23 frames) | 0 | 0 |
| Banned frames added later (25 frames) | 5 problems | 70 problems |
The second banned list was added after an error analysis on 2026-10-19 and was applied to the newer v4 and v4.1 generators. Its only hit is the phrase "two pieces of" in v2-style measurement problems written before it existed (about 0.2% of problems), e.g. "Two pieces of fence measure 95 feet and 19 feet; end to end they reach 114 feet." We left these in: the sample shows the data as it was trained on.
The generators never read benchmark rows. Benchmarks were used only to score models before and after a change; changes to the generators were judged on our own held-out phrasings.
- Downloads last month
- 29