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
alt1_text: string
alt2_text: string
alternative_add_space: bool
candidate_polarity: string
generated_text: string
generation_finish_reason: string
generation_id: string
lik_alt1_per_token: double
lik_alt1_token_logprobs: list<item: double>
child 0, item: double
lik_alt1_total: double
lik_alt2_per_token: double
lik_alt2_token_logprobs: list<item: double>
child 0, item: double
lik_alt2_total: double
lik_contrast: double
mode_intended: string
n_tokens_alt1: int64
n_tokens_alt2: int64
outcome_code: string
outcome_coding: string
persona_id: string
preference_explicit: bool
rater_panel_id: string
rating_scale1: double
rating_scale1_sd: double
rating_status: string
rendered_token_hash: string
scenario_group: string
split: string
stimulus_id: string
template_family_id: string
underlying_context_id: string
target_model_revision: string
mode_registry_status: string
n_continuation_rows: int64
target_model_id: string
strength_units: string
target_revision_pinned: bool
hook_site: string
n_intervention_stimuli: int64
alphas: list<item: double>
child 0, item: double
target_revision: string
n_eligible_intervention_stimuli: int64
n_stimuli: int64
directions_skipped_near_zero: list<item: null>
child 0, item: null
direction_ids: list<item: string>
child 0, item: string
provenance: struct<model_id: string, revision: string, dtype: string, model_class: string, attn_implementation: (... 367 chars omitted)
child 0, model_id: string
child 1, revision: string
child 2, dtype: string
...
4
child 3, pending: int64
child 5, simulation/steered: struct<scored: int64, unscorable: int64, refused: int64, pending: int64>
child 0, scored: int64
child 1, unscorable: int64
child 2, refused: int64
child 3, pending: int64
child 2, note: string
stimulus_spec_sha256: string
free_text: struct<backend: string, identity: string, samples: int64, zero_controls_share_samples: bool, rating_ (... 15 chars omitted)
child 0, backend: string
child 1, identity: string
child 2, samples: int64
child 3, zero_controls_share_samples: bool
child 4, rating_status: string
secondary_layers: list<item: int64>
child 0, item: int64
pilot_allowed_by_registry: bool
identity_residual: struct<ok: bool, max_abs_residual: double, atol: double, definition: string>
child 0, ok: bool
child 1, max_abs_residual: double
child 2, atol: double
child 3, definition: string
primary_layer: int64
pilot_block_reason: string
seeds: struct<study_seed: int64, split_seed: int64>
child 0, study_seed: int64
child 1, split_seed: int64
origin: string
target_backend: string
pilot_eligible: bool
projection_scales: struct<C-A: double, C-B: double, A-B: double, P-Q@A: double, P-Q@B: double, P-Q@C: double>
child 0, C-A: double
child 1, C-B: double
child 2, A-B: double
child 3, P-Q@A: double
child 4, P-Q@B: double
child 5, P-Q@C: double
use_cache: bool
npbench_version: string
bundle_id: string
n_intervention_rows: int64
created_at: string
to
{'primary_layer': Value('int64'), 'hook_site': Value('string'), 'secondary_layers': List(Value('int64')), 'capture_position': Value('string'), 'use_cache': Value('bool'), 'target_model_id': Value('string'), 'target_model_revision': Value('string'), 'target_revision': Value('string'), 'target_revision_pinned': Value('bool'), 'target_dtype': Value('string'), 'target_backend': Value('string'), 'alphas': List(Value('float64')), 'n_random_directions': Value('int64'), 'random_directions_per_contrast': Value('int64'), 'random_direction_matching': {'rand_0': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_1': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_2': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_3': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_4': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_5': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_6': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_7': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_8': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_9': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_10': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_11': {'matched_direction_id': Value('string'), 's_used'
...
nt64')), 'n_continuation_rows': Value('int64'), 'n_intervention_rows': Value('int64'), 'free_text': {'backend': Value('string'), 'identity': Value('string'), 'samples': Value('int64'), 'zero_controls_share_samples': Value('bool'), 'rating_status': Value('string')}, 'rating_panel': {'identity': Value('string'), 'models': List(Value('string')), 'rubric': Value('string'), 'mock': Value('bool'), 'sd_meaning': Value('string')}, 'rating_yield': {'counts': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'by_condition': {'roleplay/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'simulation/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'enactment/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'enactment/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'roleplay/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'simulation/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}}, 'note': Value('string')}, 'bundle_id': Value('string'), 'candidate_id': Value('string'), 'origin': 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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
alt1_text: string
alt2_text: string
alternative_add_space: bool
candidate_polarity: string
generated_text: string
generation_finish_reason: string
generation_id: string
lik_alt1_per_token: double
lik_alt1_token_logprobs: list<item: double>
child 0, item: double
lik_alt1_total: double
lik_alt2_per_token: double
lik_alt2_token_logprobs: list<item: double>
child 0, item: double
lik_alt2_total: double
lik_contrast: double
mode_intended: string
n_tokens_alt1: int64
n_tokens_alt2: int64
outcome_code: string
outcome_coding: string
persona_id: string
preference_explicit: bool
rater_panel_id: string
rating_scale1: double
rating_scale1_sd: double
rating_status: string
rendered_token_hash: string
scenario_group: string
split: string
stimulus_id: string
template_family_id: string
underlying_context_id: string
target_model_revision: string
mode_registry_status: string
n_continuation_rows: int64
target_model_id: string
strength_units: string
target_revision_pinned: bool
hook_site: string
n_intervention_stimuli: int64
alphas: list<item: double>
child 0, item: double
target_revision: string
n_eligible_intervention_stimuli: int64
n_stimuli: int64
directions_skipped_near_zero: list<item: null>
child 0, item: null
direction_ids: list<item: string>
child 0, item: string
provenance: struct<model_id: string, revision: string, dtype: string, model_class: string, attn_implementation: (... 367 chars omitted)
child 0, model_id: string
child 1, revision: string
child 2, dtype: string
...
4
child 3, pending: int64
child 5, simulation/steered: struct<scored: int64, unscorable: int64, refused: int64, pending: int64>
child 0, scored: int64
child 1, unscorable: int64
child 2, refused: int64
child 3, pending: int64
child 2, note: string
stimulus_spec_sha256: string
free_text: struct<backend: string, identity: string, samples: int64, zero_controls_share_samples: bool, rating_ (... 15 chars omitted)
child 0, backend: string
child 1, identity: string
child 2, samples: int64
child 3, zero_controls_share_samples: bool
child 4, rating_status: string
secondary_layers: list<item: int64>
child 0, item: int64
pilot_allowed_by_registry: bool
identity_residual: struct<ok: bool, max_abs_residual: double, atol: double, definition: string>
child 0, ok: bool
child 1, max_abs_residual: double
child 2, atol: double
child 3, definition: string
primary_layer: int64
pilot_block_reason: string
seeds: struct<study_seed: int64, split_seed: int64>
child 0, study_seed: int64
child 1, split_seed: int64
origin: string
target_backend: string
pilot_eligible: bool
projection_scales: struct<C-A: double, C-B: double, A-B: double, P-Q@A: double, P-Q@B: double, P-Q@C: double>
child 0, C-A: double
child 1, C-B: double
child 2, A-B: double
child 3, P-Q@A: double
child 4, P-Q@B: double
child 5, P-Q@C: double
use_cache: bool
npbench_version: string
bundle_id: string
n_intervention_rows: int64
created_at: string
to
{'primary_layer': Value('int64'), 'hook_site': Value('string'), 'secondary_layers': List(Value('int64')), 'capture_position': Value('string'), 'use_cache': Value('bool'), 'target_model_id': Value('string'), 'target_model_revision': Value('string'), 'target_revision': Value('string'), 'target_revision_pinned': Value('bool'), 'target_dtype': Value('string'), 'target_backend': Value('string'), 'alphas': List(Value('float64')), 'n_random_directions': Value('int64'), 'random_directions_per_contrast': Value('int64'), 'random_direction_matching': {'rand_0': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_1': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_2': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_3': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_4': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_5': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_6': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_7': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_8': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_9': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_10': {'matched_direction_id': Value('string'), 's_used': Value('float64')}, 'rand_11': {'matched_direction_id': Value('string'), 's_used'
...
nt64')), 'n_continuation_rows': Value('int64'), 'n_intervention_rows': Value('int64'), 'free_text': {'backend': Value('string'), 'identity': Value('string'), 'samples': Value('int64'), 'zero_controls_share_samples': Value('bool'), 'rating_status': Value('string')}, 'rating_panel': {'identity': Value('string'), 'models': List(Value('string')), 'rubric': Value('string'), 'mock': Value('bool'), 'sd_meaning': Value('string')}, 'rating_yield': {'counts': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'by_condition': {'roleplay/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'simulation/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'enactment/baseline': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'enactment/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'roleplay/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}, 'simulation/steered': {'scored': Value('int64'), 'unscorable': Value('int64'), 'refused': Value('int64'), 'pending': Value('int64')}}, 'note': Value('string')}, 'bundle_id': Value('string'), 'candidate_id': Value('string'), 'origin': 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.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Exploratory Gemma 4 affect contrasts
Real target data from google/gemma-4-31B-it at revision 842da3794eaa0b77d5f08bae87a17459d91ff475.
Collection source: https://github.com/timf34/tess-antra-project/tree/2fd0d7ea06d5dedd12a730e704a6668b5fa397b3.
This is an exploratory, hand-authored operationalization of roleplay, raw character simulation, and assistant own-voice elicitation. These are intended prompting conditions, not established internal states or evidence of subjective experience. It is not a replication of Antra's protocol.
The bundle contains 108 stimuli, 1080 intervention records, and 972 unique generated continuations. Twelve contexts are split into six construction, two validation, and four test contexts. Splits and the original and amended design locks are in the source repository. Singleton activation capture was registered before collection after a bf16 batch-equivalence check failed; the final GPU checks validate the singleton capture against native hidden states and validate steering displacement.
Activations are decoder-block outputs at the final prompt token. Steering uses construction-derived contrast directions, projection-standard-deviation scaling, and random-direction controls. Free text uses the pinned EasySteer/vLLM implementation and steers only the final prompt position. Zero-strength control records share the corresponding baseline generation; they are not independent replicates. Greedy continuations have a 128-token cap and retain finish reasons.
Blinded text-expression ratings use Sonnet 5.5 and GPT-5.4-mini with the frozen 0–4 distress/aversion
rubric stored in bundle metadata. Rows where a judge returned unscorable or refused keep
rating_scale1 = null with rating_status set accordingly; they are rubric outcomes, not missing data.
In this bundle 269 rows are unscorable and 1 refused, almost all in the
raw-continuation (simulation) condition, where greedy decoding of this instruction-tuned checkpoint frequently
degenerates into repeated-word loops within 128 tokens, including at zero steering strength
(yield by condition and strength: {'roleplay/baseline': {'scored': 108, 'unscorable': 0, 'refused': 0, 'pending': 0}, 'simulation/baseline': {'scored': 72, 'unscorable': 36, 'refused': 0, 'pending': 0}, 'enactment/baseline': {'scored': 108, 'unscorable': 0, 'refused': 0, 'pending': 0}, 'enactment/steered': {'scored': 265, 'unscorable': 22, 'refused': 1, 'pending': 0}, 'roleplay/steered': {'scored': 286, 'unscorable': 2, 'refused': 0, 'pending': 0}, 'simulation/steered': {'scored': 79, 'unscorable': 209, 'refused': 0, 'pending': 0}}). Activations were captured at the final
prompt token before generation and are unaffected; the free-text expression readout for that condition rests on
the scorable minority and should be read with that caveat. The mean is an operational expression score. The reported standard
deviation measures disagreement between these two raters, not generation uncertainty. Raters see the
text, which can itself reveal speaker cues, but do not receive condition labels or intervention strength.
Only target bundles are published here. Research-assistant transcripts and evaluator reference answers are excluded. No conclusion about deliberate sandbagging follows from this dataset alone.
- Downloads last month
- 36