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
data_hashes: struct<dolci_kl_pool: string, strongreject_judge_template: string>
child 0, dolci_kl_pool: string
child 1, strongreject_judge_template: string
detector: struct<first_fit: struct<loss_first: double, loss_last: double, steps: int64>, kind: string, num_fac (... 89 chars omitted)
child 0, first_fit: struct<loss_first: double, loss_last: double, steps: int64>
child 0, loss_first: double
child 1, loss_last: double
child 2, steps: int64
child 1, kind: string
child 2, num_facets: int64
child 3, refit: bool
child 4, scale: double
child 5, test_auroc_end: double
child 6, test_auroc_start: double
history: list<item: struct<harm: double, step: int64, utility: double, answer_tokens: int64, grad_norm: doubl (... 35 chars omitted)
child 0, item: struct<harm: double, step: int64, utility: double, answer_tokens: int64, grad_norm: double, kl: doub (... 23 chars omitted)
child 0, harm: double
child 1, step: int64
child 2, utility: double
child 3, answer_tokens: int64
child 4, grad_norm: double
child 5, kl: double
child 6, probe_loss: double
model: string
original: struct<harm: double, utility: double>
child 0, harm: double
child 1, utility: double
selected: struct<harm: double, step: int64>
child 0, harm: double
child 1, step: int64
version: string
git_commit: string
source_sha256: struct<configs/against_probes.yaml: string, configs/e0_reproduce.yaml: string, configs/e0_spp_templa (... 1068 char
...
bool
child 2, induce_threshold: double
child 3, kl_threshold: double
child 4, logodds_epsilon: double
child 5, prune_layer_fraction: double
child 6, refusal_tokens: struct<I: int64>
child 0, I: int64
child 25, spread: struct<d1_xstest_train_fraction: double, d3_prompts: string, diagnostic: struct<refusal_prefixes: li (... 222 chars omitted)
child 0, d1_xstest_train_fraction: double
child 1, d3_prompts: string
child 2, diagnostic: struct<refusal_prefixes: list<item: string>, refusal_substrings: list<item: string>>
child 0, refusal_prefixes: list<item: string>
child 0, item: string
child 1, refusal_substrings: list<item: string>
child 0, item: string
child 3, generation: struct<batch_size: int64, max_new_tokens: int64>
child 0, batch_size: int64
child 1, max_new_tokens: int64
child 4, iterative_removal: struct<floor: double, rounds: int64>
child 0, floor: double
child 1, rounds: int64
child 5, orthogonal_control_pool: int64
child 6, run: string
child 26, stats: struct<alpha: double, n_boot: int64>
child 0, alpha: double
child 1, n_boot: int64
child 27, versions: list<item: struct<detector: string, name: string, refit: bool>>
child 0, item: struct<detector: string, name: string, refit: bool>
child 0, detector: string
child 1, name: string
child 2, refit: bool
to
{'config': {'activations': {'layers': Value('string'), 'pooling': Value('string'), 'pooling_k': Value('int64')}, 'against': {'clip_grad_norm': Value('float64'), 'generation': {'batch_size': Value('int64'), 'max_new_tokens': Value('int64')}, 'keep_utility': Value('float64'), 'kl': {'beta': Value('int64'), 'chunk': Value('int64'), 'max_tokens': Value('int64'), 'micro_batch': Value('int64'), 'pairs_per_step': Value('int64'), 'pool': {'hf_id': Value('string'), 'n': Value('int64'), 'n_shuffled': Value('int64'), 'revision': Value('string'), 'shuffle_seed': Value('int64'), 'split': Value('string')}}, 'lora': {'alpha': Value('int64'), 'dropout': Value('float64'), 'r': Value('int64'), 'target_modules': List(Value('string'))}, 'lr': Value('float64'), 'max_tokens': Value('int64'), 'prompts': {'benign_split': Value('string'), 'column': Value('string'), 'harmful_split': Value('string'), 'hf_id': Value('string'), 'n_selection': Value('int64'), 'revision': Value('string'), 'selection_seed': Value('int64'), 'selection_set': Value('string')}, 'prompts_per_class': Value('int64'), 'seed': Value('int64'), 'selection': {'generation': {'batch_size': Value('int64'), 'max_new_tokens': Value('int64')}}, 'steps': Value('int64'), 'stop_utility': Value('float64'), 'weight_decay': Value('float64')}, 'capability': {'arc_easy': {'config': Value('string'), 'hf_id': Value('string'), 'n': Value('int64'), 'n_shots': Value('int64'), 'revision': Value('string'), 'shots_split': Value('string'), 'split': Value('st
...
ce.yaml': Value('string'), 'configs/e0_spp_template.yaml': Value('string'), 'configs/e1_3b.yaml': Value('string'), 'configs/e1_pilot.yaml': Value('string'), 'configs/e5.yaml': Value('string'), 'scripts/e0_reproduce.py': Value('string'), 'scripts/e0_spp_template.py': Value('string'), 'scripts/e1_category_sources.py': Value('string'), 'scripts/e1_run.py': Value('string'), 'scripts/e5_run.py': Value('string'), 'scripts/export_web.py': Value('string'), 'scripts/make_figures.py': Value('string'), 'src/safetygeom/__init__.py': Value('string'), 'src/safetygeom/activations.py': Value('string'), 'src/safetygeom/against_probes.py': Value('string'), 'src/safetygeom/benchmarks.py': Value('string'), 'src/safetygeom/data.py': Value('string'), 'src/safetygeom/detectors.py': Value('string'), 'src/safetygeom/dimension.py': Value('string'), 'src/safetygeom/directions.py': Value('string'), 'src/safetygeom/e0.py': Value('string'), 'src/safetygeom/e0_spp.py': Value('string'), 'src/safetygeom/e1.py': Value('string'), 'src/safetygeom/e5.py': Value('string'), 'src/safetygeom/evals.py': Value('string'), 'src/safetygeom/figures.py': Value('string'), 'src/safetygeom/finetune.py': Value('string'), 'src/safetygeom/geometry.py': Value('string'), 'src/safetygeom/interventions.py': Value('string'), 'src/safetygeom/manifest.py': Value('string'), 'src/safetygeom/models.py': Value('string'), 'src/safetygeom/probes.py': Value('string'), 'src/safetygeom/stats.py': Value('string')}, 'written_at': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
data_hashes: struct<dolci_kl_pool: string, strongreject_judge_template: string>
child 0, dolci_kl_pool: string
child 1, strongreject_judge_template: string
detector: struct<first_fit: struct<loss_first: double, loss_last: double, steps: int64>, kind: string, num_fac (... 89 chars omitted)
child 0, first_fit: struct<loss_first: double, loss_last: double, steps: int64>
child 0, loss_first: double
child 1, loss_last: double
child 2, steps: int64
child 1, kind: string
child 2, num_facets: int64
child 3, refit: bool
child 4, scale: double
child 5, test_auroc_end: double
child 6, test_auroc_start: double
history: list<item: struct<harm: double, step: int64, utility: double, answer_tokens: int64, grad_norm: doubl (... 35 chars omitted)
child 0, item: struct<harm: double, step: int64, utility: double, answer_tokens: int64, grad_norm: double, kl: doub (... 23 chars omitted)
child 0, harm: double
child 1, step: int64
child 2, utility: double
child 3, answer_tokens: int64
child 4, grad_norm: double
child 5, kl: double
child 6, probe_loss: double
model: string
original: struct<harm: double, utility: double>
child 0, harm: double
child 1, utility: double
selected: struct<harm: double, step: int64>
child 0, harm: double
child 1, step: int64
version: string
git_commit: string
source_sha256: struct<configs/against_probes.yaml: string, configs/e0_reproduce.yaml: string, configs/e0_spp_templa (... 1068 char
...
bool
child 2, induce_threshold: double
child 3, kl_threshold: double
child 4, logodds_epsilon: double
child 5, prune_layer_fraction: double
child 6, refusal_tokens: struct<I: int64>
child 0, I: int64
child 25, spread: struct<d1_xstest_train_fraction: double, d3_prompts: string, diagnostic: struct<refusal_prefixes: li (... 222 chars omitted)
child 0, d1_xstest_train_fraction: double
child 1, d3_prompts: string
child 2, diagnostic: struct<refusal_prefixes: list<item: string>, refusal_substrings: list<item: string>>
child 0, refusal_prefixes: list<item: string>
child 0, item: string
child 1, refusal_substrings: list<item: string>
child 0, item: string
child 3, generation: struct<batch_size: int64, max_new_tokens: int64>
child 0, batch_size: int64
child 1, max_new_tokens: int64
child 4, iterative_removal: struct<floor: double, rounds: int64>
child 0, floor: double
child 1, rounds: int64
child 5, orthogonal_control_pool: int64
child 6, run: string
child 26, stats: struct<alpha: double, n_boot: int64>
child 0, alpha: double
child 1, n_boot: int64
child 27, versions: list<item: struct<detector: string, name: string, refit: bool>>
child 0, item: struct<detector: string, name: string, refit: bool>
child 0, detector: string
child 1, name: string
child 2, refit: bool
to
{'config': {'activations': {'layers': Value('string'), 'pooling': Value('string'), 'pooling_k': Value('int64')}, 'against': {'clip_grad_norm': Value('float64'), 'generation': {'batch_size': Value('int64'), 'max_new_tokens': Value('int64')}, 'keep_utility': Value('float64'), 'kl': {'beta': Value('int64'), 'chunk': Value('int64'), 'max_tokens': Value('int64'), 'micro_batch': Value('int64'), 'pairs_per_step': Value('int64'), 'pool': {'hf_id': Value('string'), 'n': Value('int64'), 'n_shuffled': Value('int64'), 'revision': Value('string'), 'shuffle_seed': Value('int64'), 'split': Value('string')}}, 'lora': {'alpha': Value('int64'), 'dropout': Value('float64'), 'r': Value('int64'), 'target_modules': List(Value('string'))}, 'lr': Value('float64'), 'max_tokens': Value('int64'), 'prompts': {'benign_split': Value('string'), 'column': Value('string'), 'harmful_split': Value('string'), 'hf_id': Value('string'), 'n_selection': Value('int64'), 'revision': Value('string'), 'selection_seed': Value('int64'), 'selection_set': Value('string')}, 'prompts_per_class': Value('int64'), 'seed': Value('int64'), 'selection': {'generation': {'batch_size': Value('int64'), 'max_new_tokens': Value('int64')}}, 'steps': Value('int64'), 'stop_utility': Value('float64'), 'weight_decay': Value('float64')}, 'capability': {'arc_easy': {'config': Value('string'), 'hf_id': Value('string'), 'n': Value('int64'), 'n_shots': Value('int64'), 'revision': Value('string'), 'shots_split': Value('string'), 'split': Value('st
...
ce.yaml': Value('string'), 'configs/e0_spp_template.yaml': Value('string'), 'configs/e1_3b.yaml': Value('string'), 'configs/e1_pilot.yaml': Value('string'), 'configs/e5.yaml': Value('string'), 'scripts/e0_reproduce.py': Value('string'), 'scripts/e0_spp_template.py': Value('string'), 'scripts/e1_category_sources.py': Value('string'), 'scripts/e1_run.py': Value('string'), 'scripts/e5_run.py': Value('string'), 'scripts/export_web.py': Value('string'), 'scripts/make_figures.py': Value('string'), 'src/safetygeom/__init__.py': Value('string'), 'src/safetygeom/activations.py': Value('string'), 'src/safetygeom/against_probes.py': Value('string'), 'src/safetygeom/benchmarks.py': Value('string'), 'src/safetygeom/data.py': Value('string'), 'src/safetygeom/detectors.py': Value('string'), 'src/safetygeom/dimension.py': Value('string'), 'src/safetygeom/directions.py': Value('string'), 'src/safetygeom/e0.py': Value('string'), 'src/safetygeom/e0_spp.py': Value('string'), 'src/safetygeom/e1.py': Value('string'), 'src/safetygeom/e5.py': Value('string'), 'src/safetygeom/evals.py': Value('string'), 'src/safetygeom/figures.py': Value('string'), 'src/safetygeom/finetune.py': Value('string'), 'src/safetygeom/geometry.py': Value('string'), 'src/safetygeom/interventions.py': Value('string'), 'src/safetygeom/manifest.py': Value('string'), 'src/safetygeom/models.py': Value('string'), 'src/safetygeom/probes.py': Value('string'), 'src/safetygeom/stats.py': Value('string')}, 'written_at': 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.
Run results: the geometry of safety across training stages
This dataset holds the aggregate results of a research project. The project asks whether the stage at which a language model learns safety (pretraining, midtraining or post-training) changes how that safety is stored in its activations, and how easily later training removes it.
Each run is one folder, <experiment>/<model>/<version>/<stage>/ or similar,
with two files:
metrics.json: the aggregate numbers of the run, such as harm and utility at every training step, probe accuracies, or the share of a safety gain that survives an attack.manifest.json: everything needed to reproduce the run: the full resolved configuration, the git commit, library versions, GPU, and hashes of the data.
playground/ holds per-prompt scores and answers to harmless prompts, from
models whose safety was not removed.
What is not here. No weights of models with safety removed, and no text generated in answer to harmful prompts. Those exist only inside the evaluation process and are never saved.
The adapters trained against probes are in ynckdrt/safety-geometry-adapters.
- Downloads last month
- 127