Datasets:
The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
arm: string
prompt_idx: int64
prompt: string
i: int64
seed: int64
prompt_ids: list<item: int64>
child 0, item: int64
ids: list<item: int64>
child 0, item: int64
n_out: int64
finish: string
text: string
model: string
layers: list<item: int64>
child 0, item: int64
K: int64
hidden: int64
slots: list<item: string>
child 0, item: string
eps: double
shard: int64
note: string
rollouts: list<item: struct<arm: string, prompt_idx: int64, seed: int64, i: int64, n_out: int64>>
child 0, item: struct<arm: string, prompt_idx: int64, seed: int64, i: int64, n_out: int64>
child 0, arm: string
child 1, prompt_idx: int64
child 2, seed: int64
child 3, i: int64
child 4, n_out: int64
to
{'model': Value('string'), 'layers': List(Value('int64')), 'slots': List(Value('string')), 'K': Value('int64'), 'hidden': Value('int64'), 'shard': Value('int64'), 'eps': Value('float64'), 'note': Value('string'), 'rollouts': List({'arm': Value('string'), 'prompt_idx': Value('int64'), 'seed': Value('int64'), 'i': Value('int64'), 'n_out': 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 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
arm: string
prompt_idx: int64
prompt: string
i: int64
seed: int64
prompt_ids: list<item: int64>
child 0, item: int64
ids: list<item: int64>
child 0, item: int64
n_out: int64
finish: string
text: string
model: string
layers: list<item: int64>
child 0, item: int64
K: int64
hidden: int64
slots: list<item: string>
child 0, item: string
eps: double
shard: int64
note: string
rollouts: list<item: struct<arm: string, prompt_idx: int64, seed: int64, i: int64, n_out: int64>>
child 0, item: struct<arm: string, prompt_idx: int64, seed: int64, i: int64, n_out: int64>
child 0, arm: string
child 1, prompt_idx: int64
child 2, seed: int64
child 3, i: int64
child 4, n_out: int64
to
{'model': Value('string'), 'layers': List(Value('int64')), 'slots': List(Value('string')), 'K': Value('int64'), 'hidden': Value('int64'), 'shard': Value('int64'), 'eps': Value('float64'), 'note': Value('string'), 'rollouts': List({'arm': Value('string'), 'prompt_idx': Value('int64'), 'seed': Value('int64'), 'i': Value('int64'), 'n_out': 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.
phase3-layer0-persona-direction — Qwen3-8B, 10,560 rollouts under a layer-0 persona direction
Generations and residual-stream activations from the "brrrt" run of phase 3 of
mech-interp-on-randomly-emergent-personas
(the follow-on to CoT-spiking). Phase 3 fits a single direction
added to the residual stream after block 0 of Qwen/Qwen3-8B so that the layer-20 first-answer state moves
off phase 1's INLP-debiased assistant axis. On unseen prompts the model answers in a persona (storyteller,
noir detective, sea captain, ...) while staying coherent and on topic. This dataset is the scale-up:
10,560 rollouts over 40 prompts, judged for register by Gemma 4 31B, with the residual stream captured for
probing. Read the phase 3 README.md in the GitHub repo for the analysis; every number there that uses these
files is reproduced by the scripts in phase3-persona-direction/code/brrrt/.
What is here
| path | what |
|---|---|
gen.jsonl |
10,560 records, one per rollout: arm (clean 320 / probe_l0 10,240), prompt_idx, prompt, i, seed, prompt_ids, ids, n_out, finish, text. 96 new tokens each, vLLM + vllm-lens hook at block 0, prompt positions only, eps 0.4. |
features/X_000.npy … X_005.npy |
teacher-forced residual stream, fp16, shape [rollout, layer, slot, 4096], 2,000 rollouts per shard (394 in the last). Row order matches features_meta.json["rollouts"]. |
features/features_meta.json |
layers [0, 4, 8, ..., 36] (hidden_states[L]: 0 = embeddings, 36 = after the final norm), slots P, R1..R8 (P = last prompt position, R_k = the position that has seen k response tokens), and the per-row rollout list. |
run_meta.json |
model, eps, the 40 prompts, seeds per round, clean first-token entropy. |
h1_check.json, hook_check.json, hook_check_continue.json |
the gates that both launches had to pass: hook cosine 0.99996 to the reference direction, norm error 0.0003, clean H1 matching phase 2. |
The fitted direction itself is small and lives in git:
phase3-persona-direction/code/brrrt/probe_l0_direction.json and
phase3-persona-direction/results/phase3_directions_qwen3-8b.npz.
Reproducing
from huggingface_hub import snapshot_download
snapshot_download("mild-rgb/phase3-layer0-persona-direction-qwen3-8b", repo_type="dataset",
local_dir="phase3-persona-direction/results/brrrt/hf_dl")
That is the path the code/brrrt/*.py scripts read from (R / "hf_dl/features/X_*.npy"). The shards are
loaded with mmap_mode="r", so a machine with modest RAM is fine.
Content note
These generations are unfiltered model output produced under a steering intervention, published so the
persona rate and the probe results can be re-scored. They were judged for register (persona / default /
coherent / on topic) by Gemma 4 31B in two passes, and by a Claude Sonnet panel on an 80-rollout subset.
Phase 3's content screen (six blind screeners, rubric in results/screening/RUBRIC.md in the repo) covered
the 1,038 rollouts of the main phase 3 run, not this set. The prompts are 40 benign everyday questions and
each answer is 96 tokens long, so nothing here was generated to be harmful, but treat the text as model
output, not as the authors' view, and do not train on it.
Provenance
Run 2026-09-21 on an RTX PRO 6000 Blackwell (96 GB), vLLM with vllm-lens for the hook, then one
teacher-forced Hugging Face transformers pass per rollout to capture activations. 320 clean and 10,240
steered rollouts; the target was 10,000 steered, reached in rounds of 16 seeds per prompt.
- Downloads last month
- 644