s1lv3rj1nx/openjev-general-lora
Text Classification • Updated
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
state: struct<sentence1: string, sentence2: string>
child 0, sentence1: string
child 1, sentence2: string
answers: struct<answer: string>
child 0, answer: string
meta: struct<tier: string, source: string>
child 0, tier: string
child 1, source: string
criteria: struct<>
description: string
name: string
recall_floors: struct<>
holdout_of: list<item: null>
child 0, item: null
cost_escalate: null
questions: struct<answer: struct<instructions: string, criteria: struct<aunt: string, brother: string, daughter (... 306 chars omitted)
child 0, answer: struct<instructions: string, criteria: struct<aunt: string, brother: string, daughter: string, daugh (... 290 chars omitted)
child 0, instructions: string
child 1, criteria: struct<aunt: string, brother: string, daughter: string, daughter-in-law: string, father: string, fat (... 236 chars omitted)
child 0, aunt: string
child 1, brother: string
child 2, daughter: string
child 3, daughter-in-law: string
child 4, father: string
child 5, father-in-law: string
child 6, granddaughter: string
child 7, grandfather: string
child 8, grandmother: string
child 9, grandson: string
child 10, mother: string
child 11, mother-in-law: string
child 12, nephew: string
child 13, niece: string
child 14, sister: string
child 15, son: string
child 16, son-in-law: string
child 17, uncle: string
child 2, type: string
costs: struct<>
to
{'name': Value('string'), 'description': Value('string'), 'questions': {'answer': {'instructions': Value('string'), 'criteria': {'aunt': Value('string'), 'brother': Value('string'), 'daughter': Value('string'), 'daughter-in-law': Value('string'), 'father': Value('string'), 'father-in-law': Value('string'), 'granddaughter': Value('string'), 'grandfather': Value('string'), 'grandmother': Value('string'), 'grandson': Value('string'), 'mother': Value('string'), 'mother-in-law': Value('string'), 'nephew': Value('string'), 'niece': Value('string'), 'sister': Value('string'), 'son': Value('string'), 'son-in-law': Value('string'), 'uncle': Value('string')}, 'type': Value('string')}}, 'costs': {}, 'cost_escalate': Value('null'), 'recall_floors': {}, 'holdout_of': List(Value('null'))}
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
state: struct<sentence1: string, sentence2: string>
child 0, sentence1: string
child 1, sentence2: string
answers: struct<answer: string>
child 0, answer: string
meta: struct<tier: string, source: string>
child 0, tier: string
child 1, source: string
criteria: struct<>
description: string
name: string
recall_floors: struct<>
holdout_of: list<item: null>
child 0, item: null
cost_escalate: null
questions: struct<answer: struct<instructions: string, criteria: struct<aunt: string, brother: string, daughter (... 306 chars omitted)
child 0, answer: struct<instructions: string, criteria: struct<aunt: string, brother: string, daughter: string, daugh (... 290 chars omitted)
child 0, instructions: string
child 1, criteria: struct<aunt: string, brother: string, daughter: string, daughter-in-law: string, father: string, fat (... 236 chars omitted)
child 0, aunt: string
child 1, brother: string
child 2, daughter: string
child 3, daughter-in-law: string
child 4, father: string
child 5, father-in-law: string
child 6, granddaughter: string
child 7, grandfather: string
child 8, grandmother: string
child 9, grandson: string
child 10, mother: string
child 11, mother-in-law: string
child 12, nephew: string
child 13, niece: string
child 14, sister: string
child 15, son: string
child 16, son-in-law: string
child 17, uncle: string
child 2, type: string
costs: struct<>
to
{'name': Value('string'), 'description': Value('string'), 'questions': {'answer': {'instructions': Value('string'), 'criteria': {'aunt': Value('string'), 'brother': Value('string'), 'daughter': Value('string'), 'daughter-in-law': Value('string'), 'father': Value('string'), 'father-in-law': Value('string'), 'granddaughter': Value('string'), 'grandfather': Value('string'), 'grandmother': Value('string'), 'grandson': Value('string'), 'mother': Value('string'), 'mother-in-law': Value('string'), 'nephew': Value('string'), 'niece': Value('string'), 'sister': Value('string'), 'son': Value('string'), 'son-in-law': Value('string'), 'uncle': Value('string')}, 'type': Value('string')}}, 'costs': {}, 'cost_escalate': Value('null'), 'recall_floors': {}, 'holdout_of': List(Value('null'))}
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.
OpenJev training mixture
279 classification and multiple-choice tasks, 323,466 rows, normalised into
one typed-decision format so a single model can be trained across all of
them. Assembled from tasksource plus
two curated ordinal datasets.
Built for OpenJev. The format is plain JSONL, so nothing here requires that code.
Composition
| primitive | tasks | what it is |
|---|---|---|
choice |
234 | pick one of K options |
noul |
34 | yes/no |
score |
11 | one level on an ordered scale |
Option counts run from 2 to 174. 80 of the tasks carry per-example criteria, meaning each row brings its own answer menu rather than sharing a fixed one, which is what the MultipleChoice family needs.
What this is disjoint from
Every task was checked against the
held-out suite
before being written, matching on full names, owner-stripped basenames and
glob patterns rather than exact strings. 23 tasks were excluded for
overlapping it, including several civil_comments configs and banking77
under three different names.
A mixture that quietly contains your evaluation data produces excellent numbers that mean nothing, so this is enforced in code rather than promised.
Cleaning applied, and why
Contradictory labels removed, 1,994 rows. The same state carrying
different answers under the same menu cannot all be correct, so the
gradient is noise. ethics_deontology alone contributed 195.
Oversized menus removed, 1,787 rows. Some MultipleChoice options are multi-paragraph passages, and a menu alone can exceed a 2,048-token budget.
Option order shuffled. tasksource ships MultipleChoice
correct-answer-first: measured, the gold sat at index 0 in 100% of rows
across all 83 ingested tasks. Stored that way it teaches "pick the first
option" to anything reading the menu as written. Shuffled at ingestion so
the data on disk is honest.
Degenerate tasks dropped. Any task whose gold is one constant string, or still lands at a fixed index after shuffling.
Verify with scripts/audit_data.py, which reports zero errors on this
mixture, and scripts/check_packing.py, which confirms all 323,466 rows
pack inside a 2,048-token budget.
Known limitations
Lopsided toward choice. 234 against 34 and 11. Models trained here
transfer well on choice and poorly on the other two, and that is a
property of this mixture rather than of any architecture. Ordinal data in
particular is scarce: most score tasks are three-level sentiment scales.
19 tasks have a class with under 10 examples, which is learned as "never predict this" rather than learned at all.
Not large enough to be finished. The Flan work suggests gain keeps accruing past ~282 tasks. This is 279, and adding more diverse tasks, especially ordinal and yes/no ones, is the most useful contribution anyone could make.
Format
Each task is a directory with task.json and train.jsonl. Specified in
docs/dataset-format.md.
{"state": "the text", "answers": {"q": "gold label"}, "meta": {"source": "tasksource id"}}
Per-example menus add a criteria field carrying that row's options.
Licence
Apache 2.0 for the assembly. Underlying datasets keep their own licences;
each task names its source in description.
- Downloads last month
- 60