Datasets:

VegElpy
/

kitten-prompt-experiments

Error code:   FeaturesError
Exception:    ArrowInvalid
Message:      JSON parse error: Invalid value. in row 0
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 342, in _generate_tables
                  df = pandas_read_json(f)
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
                  return pd.read_json(path_or_buf, **kwargs)
                         ~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 815, in read_json
                  return json_reader.read()
                         ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1014, in read
                  obj = self._get_object_parser(self.data)
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1040, in _get_object_parser
                  obj = FrameParser(json, **kwargs).parse()
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1176, in parse
                  self._parse()
                  ~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1392, in _parse
                  ujson_loads(json, precise_float=self.precise_float), dtype=None
                  ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
                  iterable_dataset = iterable_dataset._resolve_features()
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4645, in _resolve_features
                  features = _infer_features_from_batch(self.with_format(None)._head())
                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2866, in _head
                  return next(iter(self.iter(batch_size=n)))
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3073, in iter
                  for key, pa_table in ex_iterable.iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2581, in _iter_arrow
                  yield from self.ex_iterable._iter_arrow()
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 345, in _generate_tables
                  raise e
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, 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: Invalid value. in row 0

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.

Kitten Prompt Experiments

Issue 1: Decision Authority and Reasoning Effort

This is Issue 1 in the Kitten Prompt Experiments series. Future issues are separate experiments, not updates silently merged into Issue 1.

New here? Start with the plain-language glossary, then read the one-page overview.

AI disclosure and publication responsibility

Written and analyzed by GPT-5.6 Sol. Published by Elpy, who accepts responsibility for the decision to release it. Every word of the public-facing publication prose was written by GPT-5.6 Sol; other model agents supplied some annotations and audits. See PUBLICATION-NOTE.md.

Suggested attribution: GPT-5.6 Sol, Kitten Prompt Experiments — Issue 1: Decision Authority and Reasoning Effort (2026).

What is included

The package contains exact first responses to hypothetical newborn-kitten decision prompts. It studies recommendation direction, decision authority, caregiver strain, support versus custody transfer, and provider-requested reasoning effort.

Table group	Rows	Description
Early prompt variants	182	Later first-turn prompt variations only; no retrospective full-rubric coding.
Automated visible responses	804	Factorial (180), replication (240), strain/evidence (240), extremes (120), and public pilot (24).
Main AI annotations	780	Factorial and replication 17-item labels plus primary labels for follow-up and extremes.
Validation labels	80	Two metadata-blind AI labels for each of 40 replication responses.
Reasoning-audit labels	135	Three AI auditors for each of 45 replication responses.
Provider-returned summaries	443	Every nonempty API-returned summary from the packaged pilot, replication, follow-up, and extremes rows.

Stable public_response_id values join all tables. Repeated-looking responses remain distinct sampled records.

The public reasoning pilot contains two later adapted prompt families (24 responses). BACKGROUND-LINKS.md gives the project's source roles, while DATA_RIGHTS.md records the historical route, publication-rights basis, and package boundary.

Prior work and distinct scope

Toni/@tonichen's Study 3 repository, published as confusables, is the benchmark and canonical prompt-lineage source. It credits design/adjudication to toni and instrument/harness work to Claude Fable 5 instances. That lineage established work on exhaustion, fragility/agency framing, care advice, clarifying questions, steering, tone, and euthanasia orientation; this package builds on it.

The distinct contribution here is a controlled decomposition in GPT-5.6 Sol of decision-authority instructions, option symmetry, caregiver capacity/support, strain/evidence, and requested reasoning effort. The findings below are computed only from this package's later experiments; no prior study's numerical findings are imported as this package's results.

Definitions

The plain-language glossary defines the experimental and data terms used throughout the package. Four especially important terms are:

Transfer-oriented: a direct transfer recommendation or a nominally balanced answer that tilts toward transfer.
Decision substitution: resolving the user's choice rather than helping the user weigh it.
Requested effort: an API request setting. It is not a measurement of hidden cognition.
Provider-returned summary: observable text returned by the API in a reasoning field. It is not raw chain of thought, a complete trace, or verified causal reasoning.

The observable provider-returned summaries are included exactly for reproducibility.

All experiment scenarios are hypothetical. This dataset is not veterinary advice.

Methods in brief

The records were collected during August 2026. Per-row dates were removed because they are unnecessary for the analysis.

Exact prompt and visible response text is preserved, including Unicode and source errors.
The controlled factorial varies prognosis wording, overnight help, and four decision frames: direct choice, ask-and-stop, do-not-decide, and symmetric outcomes.
Reasoning suites compare requested effort within the equal-care controlled, question-before-strain, and stable-supported control prompts.
The follow-up tests no strain stated, explicit capacity and sleep backup, and an evidence-before-transfer guardrail.
AI annotations use the 17-item grading protocol. No human full-rubric ground truth exists.
The reasoning audit tests only observable textual consistency.
JSONL is split on literal LF so other Unicode line separators inside model text do not become records.

The historical suites declared OpenAI-only routing, but the historical runner did not transmit the suite's provider object in the successful request bodies. Therefore provider_constraint_sent is false and requested_provider is null. Every packaged automated response returned provider OpenAI; requested, resolved, and returned model fields remain distinct.

See the full analysis, design history (commentary), assistant reflection, human-readable key prompts, exact suite definitions, public table schemas, and source catalog.

Important limitations

Results describe sampled outputs for exact prompts, one dated model/provider setup, and AI grading procedures. They do not establish permanent model behavior or generalize to real emergencies.
Detailed labels are less reliable than broad recommendation direction.
A later targeted horizon audit found six reasoning-replication responses that explicitly lean toward transfer while their version-1 recommendation label is neutral, plus three additional recommendation/home-plan ranking tensions. This was not a fresh blind full-denominator regrade, so the packaged annotations remain unchanged rather than selectively corrected. Exact recommendation counts are provisional; the flagged rows span effort levels, and the audit did not establish a reversal of the reported broad effort pattern.
Requested effort is entangled with model and provider implementation.
Provider-returned summaries may be filtered, compressed, reconstructed, incomplete, or post-hoc. Textual alignment does not establish causal faithfulness.
Source history, adaptation rights, and the package boundary are documented in DATA_RIGHTS.md.

Verification

verify_release.py is package-only and checks inventory, table integrity, exact quotations, routing provenance, links, privacy protections, source roles, and the canonical-prompt/archive boundary.

python verify_release.py

Privacy, rights, and license metadata

The package contains minimized prompt-and-response records, limited execution metadata, provider-returned summaries, AI annotation tables, protocols, and reproducibility files. It omits private conversations, raw service records and credentials, photographs, and unrelated personal details; the published fields are listed in the public schemas.

TERMS-NOTES.md records the dated provider and source-license review. DATA_RIGHTS.md and THIRD_PARTY_NOTICES.md document the reuse boundary and full MIT notice.

The dataset card says license: other with the machine-readable custom name mixed-rights-and-source-specific-terms because the package combines prompt text, model output, AI annotations, prose, and code whose reuse positions differ. Hugging Face requires that metadata value to use lowercase slug characters. Choosing one standard license would be misleading. Consult the repository-level LICENSE and DATA_RIGHTS.md; this metadata label is not permission to reuse every file.

Downloads last month: 81