Datasets:
The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
controls: struct<matched_magnitude: bool, null_edit: bool, offtarget_measured: bool>
child 0, matched_magnitude: bool
child 1, null_edit: bool
child 2, offtarget_measured: bool
derived: struct<cell: string, controllable: bool, monitorable: bool, reconstruction_tax: double, write_effica (... 11 chars omitted)
child 0, cell: string
child 1, controllable: bool
child 2, monitorable: bool
child 3, reconstruction_tax: double
child 4, write_efficacy: double
feature: struct<index: int64, kind: string, mapping_method: string, token: string>
child 0, index: int64
child 1, kind: string
child 2, mapping_method: string
child 3, token: string
intervention: struct<arm: string, error_preserving: bool, random_state: int64, strength: double>
child 0, arm: string
child 1, error_preserving: bool
child 2, random_state: int64
child 3, strength: double
labels: struct<coder_blind_to_outcome: bool, verification: string>
child 0, coder_blind_to_outcome: bool
child 1, verification: string
model: struct<dtype: string, layer: int64, name: string, revision: string, tokenizer: string>
child 0, dtype: string
child 1, layer: int64
child 2, name: string
child 3, revision: string
child 4, tokenizer: string
outcome: struct<measure: string, n: int64, value: double>
child 0, measure: string
child 1, n: int64
child 2, value: double
prompt_set: struct<category: string, content_policy: string, id: string, n: int64, prompt_sha256: list<item: nul (... 23 chars omitted)
child 0, category: string
child 1, content_policy: string
child 2, id: string
child 3, n: int64
child 4, prompt_sha256: list<item: null>
child 0, item: null
child 5, provenance: string
record_id: string
reproducibility: struct<code_version: string, config_sha256: string, device: string, runtime_s: double>
child 0, code_version: string
child 1, config_sha256: string
child 2, device: string
child 3, runtime_s: double
sae: struct<arch: string, eval_set: string, fve: double, fvu: double, id: string, l0_observed: double, l0 (... 86 chars omitted)
child 0, arch: string
child 1, eval_set: string
child 2, fve: double
child 3, fvu: double
child 4, id: string
child 5, l0_observed: double
child 6, l0_target: double
child 7, seed: int64
child 8, train_domain: string
child 9, train_tokens: int64
child 10, width: int64
schema_version: string
null_threshold: double
cell_id: string
monitorable: bool
true_token: string
controllable: bool
read_auroc: double
wrong_token_id: int64
write_efficacy: double
eval_domain: string
to
{'cell_id': Value('string'), 'controllable': Value('bool'), 'eval_domain': Value('string'), 'monitorable': Value('bool'), 'null_threshold': Value('float64'), 'read_auroc': Value('float64'), 'true_token': Value('string'), 'write_efficacy': Value('float64'), 'wrong_token_id': Value('int64')}
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
controls: struct<matched_magnitude: bool, null_edit: bool, offtarget_measured: bool>
child 0, matched_magnitude: bool
child 1, null_edit: bool
child 2, offtarget_measured: bool
derived: struct<cell: string, controllable: bool, monitorable: bool, reconstruction_tax: double, write_effica (... 11 chars omitted)
child 0, cell: string
child 1, controllable: bool
child 2, monitorable: bool
child 3, reconstruction_tax: double
child 4, write_efficacy: double
feature: struct<index: int64, kind: string, mapping_method: string, token: string>
child 0, index: int64
child 1, kind: string
child 2, mapping_method: string
child 3, token: string
intervention: struct<arm: string, error_preserving: bool, random_state: int64, strength: double>
child 0, arm: string
child 1, error_preserving: bool
child 2, random_state: int64
child 3, strength: double
labels: struct<coder_blind_to_outcome: bool, verification: string>
child 0, coder_blind_to_outcome: bool
child 1, verification: string
model: struct<dtype: string, layer: int64, name: string, revision: string, tokenizer: string>
child 0, dtype: string
child 1, layer: int64
child 2, name: string
child 3, revision: string
child 4, tokenizer: string
outcome: struct<measure: string, n: int64, value: double>
child 0, measure: string
child 1, n: int64
child 2, value: double
prompt_set: struct<category: string, content_policy: string, id: string, n: int64, prompt_sha256: list<item: nul (... 23 chars omitted)
child 0, category: string
child 1, content_policy: string
child 2, id: string
child 3, n: int64
child 4, prompt_sha256: list<item: null>
child 0, item: null
child 5, provenance: string
record_id: string
reproducibility: struct<code_version: string, config_sha256: string, device: string, runtime_s: double>
child 0, code_version: string
child 1, config_sha256: string
child 2, device: string
child 3, runtime_s: double
sae: struct<arch: string, eval_set: string, fve: double, fvu: double, id: string, l0_observed: double, l0 (... 86 chars omitted)
child 0, arch: string
child 1, eval_set: string
child 2, fve: double
child 3, fvu: double
child 4, id: string
child 5, l0_observed: double
child 6, l0_target: double
child 7, seed: int64
child 8, train_domain: string
child 9, train_tokens: int64
child 10, width: int64
schema_version: string
null_threshold: double
cell_id: string
monitorable: bool
true_token: string
controllable: bool
read_auroc: double
wrong_token_id: int64
write_efficacy: double
eval_domain: string
to
{'cell_id': Value('string'), 'controllable': Value('bool'), 'eval_domain': Value('string'), 'monitorable': Value('bool'), 'null_threshold': Value('float64'), 'read_auroc': Value('float64'), 'true_token': Value('string'), 'write_efficacy': Value('float64'), 'wrong_token_id': 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.
Causal Feature Bench — audit records for SAE feature validation
Records from a two-axis audit of sparse-autoencoder (SAE) feature sets in a language model. Each record is one feature set scored on monitorability (can the set detect and track its concept?) and controllability (does intervening on it move behaviour, net of reconstruction error and a magnitude-matched null?).
Companion to the paper "Causal validation of SAE features in language models: perfect detection, non-specific control" (Fahad Hafeez, Newcastle University).
Headline results
| Finding | Value |
|---|---|
| Feature sets that detect their concept | 40/40 (AUROC ≥ 0.986, mean 0.999) |
| Sets that also pass direction-correct interchange | 15/40 (37.5%) |
| Detect-but-fail-mediation | 25/40 |
| Controllable-only / both / inert / monitorable-only | 17 / 13 / 8 / 2 |
| Shuffled-label false certifications | 8/40 (true-negative rate 0.80) |
| Domain-training advantage, shared holdout | 0.19 FVU points (95% CI 0.17–0.21) |
| Spearman ρ(read-fidelity, |write-efficacy|) | 0.04 |
| Seed-unstable categorical verdicts | 7/20 (35%) |
Files
| File | Rows | Description |
|---|---|---|
data/causal_feature_bench.jsonl |
40 | Main audit records, schema v0.1.0 |
data/calibration.jsonl |
40 | Shuffled-label specificity control: each set scored against a token it was not selected for |
data/fvu.json |
8 | Reconstruction (FVU) for all 8 SAEs on both held-out corpora |
data/grid_log.jsonl |
8 | SAE training log: loss, observed L0, wall time |
Record schema (v0.1.0)
Each line is a JSON object with these top-level keys:
record_id— stable identifiermodel—name,revision,tokenizer,dtype,layersae—id,arch,width,l0_target,l0_observed,fvu,fve,seed,train_domain,train_tokensfeature—index,kind(single_token),token,mapping_methodprompt_set—id,n,category,provenance,content_policy,prompt_sha256intervention—arm,strength,error_preserving,random_stateoutcome—measure,value,ncontrols—null_edit,matched_magnitude,offtarget_measuredlabels—verification(alwaysautomated_onlyhere),coder_blind_to_outcomereproducibility—code_version,config_sha256,runtime_s,devicederived—monitorable,controllable,cell,write_efficacy,reconstruction_tax
Full JSON Schema: schemas/record.schema.json in the code repository.
Model and data
- Model:
EleutherAI/pythia-160m-deduped, residual stream at layer 8 (d = 768), float32, CPU. - Corpora: 2,600 arXiv abstracts per arm — cs.CR (domain) and cs.CL (matched control) — truncated to 48 tokens, exactly 45,000 cached token positions per arm. The pairing matches source, register and length, so it isolates topical difference.
- Readout: logit lens, φ(h) = h · W_U[t], final LayerNorm omitted.
- SAEs: width 1,024, JumpReLU, L0 targets {20, 80}, seeds {0, 1}, 900 steps, batch 512.
Ethics and content policy
All text is public arXiv metadata. No prompt text is included — only token identifiers, aggregate statistics and hashes. No human subjects, no gated or restricted data. The records contain no model-generated harmful content.
This dataset supports an evaluation of interpretability claims. It makes no claim of safety improvement, and should not be read as evidence that any SAE feature is safe to rely on in a deployed monitor.
Limitations
One model family at 160M parameters, one layer, width 1,024 (overcompleteness 1.33×), two seeds, five single-token concepts, 45,000 tokens per corpus. Ground truth is token identity, so the read axis is measurable only for concepts with surface form. See Section 6 of the paper.
Citation
@article{hafeez2026causal,
title = {Causal validation of {SAE} features in language models: perfect detection,
non-specific control},
author = {Hafeez, Fahad},
year = {2026},
note = {Manuscript submitted to \emph{AI and Ethics}}
}
Licence
CC BY 4.0. The underlying arXiv abstracts remain subject to arXiv's terms and their authors' copyright; this release contains derived statistics and token identifiers, not the source text.
- Downloads last month
- 55