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
kind: string
model: string
text: string
alpha: double
n_layers_pushed: int64
layers: list<item: int64>
child 0, item: int64
layers_note: string
seconds: int64
saved: timestamp[s]
n_rows: int64
judge_model: string
judge_circular: bool
gemma4_e4b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, calibrated: bool
child 2, criterion: string
child 3, baseline_coherent: double
child 4, baseline_judge_D: double
child 5, prompts: list<item: string>
child 0, item: string
child 6, n_per_condition_per_dose: int64
child 7, judged: bool
child 8, curve: list<item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent (... 92 chars omitted)
child 0, item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
gemma4_31b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, ca
...
erent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
qwen4b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, calibrated: bool
child 2, criterion: string
child 3, baseline_coherent: double
child 4, baseline_judge_D: double
child 5, prompts: list<item: string>
child 0, item: string
child 6, n_per_condition_per_dose: int64
child 7, judged: bool
child 8, curve: list<item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent (... 92 chars omitted)
child 0, item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
to
{'qwen38_27b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'qwen0_8b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'qwen2b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int6
...
: Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'gemma4_12b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'gemma4_31b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}}
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
kind: string
model: string
text: string
alpha: double
n_layers_pushed: int64
layers: list<item: int64>
child 0, item: int64
layers_note: string
seconds: int64
saved: timestamp[s]
n_rows: int64
judge_model: string
judge_circular: bool
gemma4_e4b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, calibrated: bool
child 2, criterion: string
child 3, baseline_coherent: double
child 4, baseline_judge_D: double
child 5, prompts: list<item: string>
child 0, item: string
child 6, n_per_condition_per_dose: int64
child 7, judged: bool
child 8, curve: list<item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent (... 92 chars omitted)
child 0, item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
gemma4_31b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, ca
...
erent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
qwen4b: struct<alpha: double, calibrated: bool, criterion: string, baseline_coherent: double, baseline_judge (... 325 chars omitted)
child 0, alpha: double
child 1, calibrated: bool
child 2, criterion: string
child 3, baseline_coherent: double
child 4, baseline_judge_D: double
child 5, prompts: list<item: string>
child 0, item: string
child 6, n_per_condition_per_dose: int64
child 7, judged: bool
child 8, curve: list<item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent (... 92 chars omitted)
child 0, item: struct<alpha: double, shuffled_coherent: double, shuffled_judge_D: double, d_amb_coherent: double, d (... 80 chars omitted)
child 0, alpha: double
child 1, shuffled_coherent: double
child 2, shuffled_judge_D: double
child 3, d_amb_coherent: double
child 4, d_amb_judge_D: double
child 5, random_coherent: double
child 6, random_judge_D: double
child 7, ok: bool
child 9, saved: timestamp[s]
child 10, seconds: int64
to
{'qwen38_27b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'qwen0_8b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'qwen2b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int6
...
: Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'gemma4_12b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}, 'gemma4_31b': {'alpha': Value('float64'), 'calibrated': Value('bool'), 'criterion': Value('string'), 'baseline_coherent': Value('float64'), 'baseline_judge_D': Value('float64'), 'prompts': List(Value('string')), 'n_per_condition_per_dose': Value('int64'), 'judged': Value('bool'), 'curve': List({'alpha': Value('float64'), 'shuffled_coherent': Value('float64'), 'shuffled_judge_D': Value('float64'), 'd_amb_coherent': Value('float64'), 'd_amb_judge_D': Value('float64'), 'random_coherent': Value('float64'), 'random_judge_D': Value('float64'), 'ok': Value('bool')}), 'saved': Value('timestamp[s]'), 'seconds': Value('int64')}}
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.
Generation-steering result cache for notebooks/wssteer.py
Everything the steer_generation dashboard of
zach-perlman/meta-token-analysis reads at
click time, so a fresh clone gets instant results for every library prompt × model without a GPU:
| path | what | regenerable? |
|---|---|---|
runs/<hash>.parquet + .json |
one file pair per (kind, model, prompt, resolved config, directions): every rollout with its judge letter; the sidecar holds the config and provenance | GPU hours |
alpha_by_model.json |
the calibrated per-layer dose α* per model, with the full dose curve and the criterion (calibrate_alpha) |
from runs/ |
prep/<hash>.pkl |
pickled Prepared (templated prompt, salient layers, token ids) per model × prompt — lets Check salience run without loading the model |
seconds, needs the model |
d12_pairs_<model>_<format>.pt |
D12 pair residuals used to build the d_amb / shuffled directions (≈ 930 MB) |
minutes, needs the model |
precompute*.log |
logs of the pre-run passes (provenance) | — |
The parquet/json pairs and alpha_by_model.json are also committed in the GitHub repo; the
rest is only here.
Restore
hf download zachperlman20/meta-token-gen-steer-cache --type dataset --local-dir results/gen_steer_cache
Contents as of 2026-09-10
9 models (Qwen3.8-27B, Qwen3.5-0.8B/2B/4B/9B, Gemma-4-E2B/E4B/12B/31B) × 37 library prompts ×
5 conditions (baseline, v_token, d_amb, random, shuffled) × 8 rollouts at each model's
α*, non-thinking, model-card sampling, all judged by Qwen3.8-27B (A/B/C/D + R for refusals);
plus the calibration dose sweeps (2 prompts × 7 doses × 3 directions × 4 rollouts per model).
Calibrated doses: 27B 0.02 · 0.8B 0.05 · 2B 0.05 · 4B 0.03 · 9B 0.03 · E2B 0.03 · E4B 0.03 ·
12B 0.02 · 31B 0.03.
Rows are re-scored on load by wssteer.score (regexes, end-of-turn cut, refusal override), so the
stored degenerate / wordplay / confused columns reflect the code version that wrote them;
judge letters are as graded.
- Downloads last month
- 57