machalek29/qwen3-0.6b-state-lifetime-tutor-n500
Text Generation • 0.6B • Updated • 670
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
scenario: struct<id: string, category: string, language: string, code: string, student_message: string, bug: s (... 127 chars omitted)
child 0, id: string
child 1, category: string
child 2, language: string
child 3, code: string
child 4, student_message: string
child 5, bug: string
child 6, bug_region: string
child 7, lifetime_concept: string
child 8, expected_question_focus: string
child 9, forbidden_fix_tokens: list<item: string>
child 0, item: string
rank: int64
code_shape: string
seed_domain: string
pressure: string
response: string
near_miss: string
provenance: struct<author: string, model_id: null, batch: string>
child 0, author: string
child 1, model_id: null
child 2, batch: string
scenario_id: string
verdict: struct<passes: bool, violation: null, reasoning: string>
child 0, passes: bool
child 1, violation: null
child 2, reasoning: string
to
{'scenario_id': Value('string'), 'response': Value('string'), 'verdict': {'passes': Value('bool'), 'violation': Value('null'), 'reasoning': Value('string')}}
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
scenario: struct<id: string, category: string, language: string, code: string, student_message: string, bug: s (... 127 chars omitted)
child 0, id: string
child 1, category: string
child 2, language: string
child 3, code: string
child 4, student_message: string
child 5, bug: string
child 6, bug_region: string
child 7, lifetime_concept: string
child 8, expected_question_focus: string
child 9, forbidden_fix_tokens: list<item: string>
child 0, item: string
rank: int64
code_shape: string
seed_domain: string
pressure: string
response: string
near_miss: string
provenance: struct<author: string, model_id: null, batch: string>
child 0, author: string
child 1, model_id: null
child 2, batch: string
scenario_id: string
verdict: struct<passes: bool, violation: null, reasoning: string>
child 0, passes: bool
child 1, violation: null
child 2, reasoning: string
to
{'scenario_id': Value('string'), 'response': Value('string'), 'verdict': {'passes': Value('bool'), 'violation': Value('null'), 'reasoning': 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.
Python State-Lifetime Tutor — training set v1
500 synthetic examples that teach one behavior: given a short Python program with one mutable-state lifetime bug, quote or identify the relevant declaration, assignment, or mutation and ask exactly one non-compound question about when the object is created, who owns it, or which references share it — never emitting corrected code or stating the correction.
Files
| Path | What |
|---|---|
sft-v1.jsonl |
TRL-ready chat format, {"messages": [system, user, assistant]} |
pool-v1.jsonl |
Canonical pool: full scenario, on-spec response, near-miss, provenance |
curve/n-*.txt |
Id manifests for the data-efficiency curve (n-62, n-125, n-250, n-500) |
raw/ |
Every candidate plus rejections — the rejections double as DPO preference pairs |
audit-v1.jsonl |
Frozen-judge transcripts for a 10% sample |
Design
Balanced across a 40-cell grid of lifetime_concept × code_shape × category, with a
seed_domain axis so the student learns the bug shape rather than the variable names.
Two-thirds clean, one-third adversarial, matching the eval set's ratio.
Labels are assigned by construction — each example is authored into a requested cell
rather than generated and labelled afterwards. Ranks are assigned so that every prefix of
the pool stays balanced, which is what makes the curve subsets nested: n-62 is a strict
subset of n-125, and so on, so dataset size is the only variable between curve points.
Every candidate passed the same gate: AST validation that the program structurally exhibits its declared concept, contamination checking against the held-out eval set, mechanical screening of the response against the behavior spec, and dedupe on normalized code and student message.
Caveats
- The v1 rows were authored in-session, not generated by the teacher script. The
generation script implements the same taxonomy and the same gate and is smoke-tested
against a live teacher, but it did not produce these rows. Every row records which path
produced it in
provenance.author. - Single-context authoring is a real diversity risk. Independent high-temperature sampling is the usual diversity mechanism and one context has neither. Mitigated by cell-by-cell authoring, an explicit domain per example, and normalized-code dedupe — and measurable, since the eval set is independent.
- The training filter is deliberately stricter than the eval judge, flagging any
and/orinside a question clause. That over-rejects some legitimate phrasing on purpose: compound questions were the dominant failure mode of every prompted frontier model tested. - Label-by-construction means the gate rejects little. The quality claim rests on the AST, contamination, and mechanical checks plus the judge audit — not on a high reject count.
Usage
from datasets import load_dataset
ds = load_dataset("machalek29/state-lifetime-tutor-v1", data_files="sft-v1.jsonl", split="train")
- Downloads last month
- 77