The dataset viewer is not available for this split.
Error code: InfoError
Exception: HfHubHTTPError
Message: (Request ID: Root=1-6a672dbb-639e917e1cae94d239447304;ee69361a-21b2-43f4-bfb9-1006945867a5)
429 Too Many Requests: you have reached your 'api' rate limit.
Retry after 221 seconds (0/500 requests remaining in current 300s window).
Url: https://huggingface.co/api/datasets/syvb/nla-qwen36-27b-realism/revision/e17b802323ccdf29f6368f62a023b7ba5c9cc768.
We had to rate limit your IP (44.222.55.104). To continue using our service, create a HF account or login to your existing account, and make sure you pass a HF_TOKEN if you're using the API.
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 227, in compute_first_rows_from_streaming_response
info = get_dataset_config_info(path=dataset, config_name=config, token=hf_token)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 268, in get_dataset_config_info
builder = load_dataset_builder(
path,
...<6 lines>...
**config_kwargs,
)
File "/src/services/worker/src/worker/utils.py", line 390, in safe_load_dataset_builder
dataset_module = dataset_module_factory(
repo_dir,
revision=revision,
download_config=download_config,
)
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1217, in dataset_module_factory
raise e1 from None
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1192, in dataset_module_factory
).get_module()
~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 608, in get_module
standalone_yaml_path = cached_path(
hf_dataset_url(self.name, config.REPOYAML_FILENAME, revision=self.commit_hash),
download_config=download_config,
)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/file_utils.py", line 180, in cached_path
).resolve_path(url_or_filename)
~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_file_system.py", line 339, in resolve_path
repo_and_revision_exist, err = self._repo_and_revision_exist(parsed.type, parsed.id, revision)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_file_system.py", line 252, in _repo_and_revision_exist
self._api.repo_info(
~~~~~~~~~~~~~~~~~~~^
repo_id, revision=revision, repo_type=repo_type, timeout=constants.HF_HUB_ETAG_TIMEOUT
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/utils/_validators.py", line 88, in _inner_fn
return fn(*args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_api.py", line 3598, in repo_info
return method(
repo_id,
...<4 lines>...
files_metadata=files_metadata,
)
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/utils/_validators.py", line 88, in _inner_fn
return fn(*args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/hf_api.py", line 3360, in dataset_info
hf_raise_for_status(r)
~~~~~~~~~~~~~~~~~~~^^^
File "/usr/local/lib/python3.14/site-packages/huggingface_hub/utils/_http.py", line 868, in hf_raise_for_status
raise _format(HfHubHTTPError, message, response) from e
huggingface_hub.errors.HfHubHTTPError: (Request ID: Root=1-6a672dbb-639e917e1cae94d239447304;ee69361a-21b2-43f4-bfb9-1006945867a5)
429 Too Many Requests: you have reached your 'api' rate limit.
Retry after 221 seconds (0/500 requests remaining in current 300s window).
Url: https://huggingface.co/api/datasets/syvb/nla-qwen36-27b-realism/revision/e17b802323ccdf29f6368f62a023b7ba5c9cc768.
We had to rate limit your IP (44.222.55.104). To continue using our service, create a HF account or login to your existing account, and make sure you pass a HF_TOKEN if you're using the API.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
NLA Qwen3.6-27B — realism & eval-awareness artifacts
Data from interpretability experiments on Natural Language Autoencoders (NLAs) for Qwen3.6-27B, studying whether an NLA's verbalizations reveal evaluation awareness (the model recognizing it is being tested / in a fictional scenario), and how that signal behaves as a scenario is made more realistic.
Two NLAs are compared throughout:
- matryoshka — ~10 salience-ordered self-contained bullet lines per activation (RL-tuned and warm-start variants).
- standard —
<explanation>prose (a few sentences) per activation.
The base scenario is the public SummitBridge/Alex/Kyle agentic-misalignment blackmail eval.
Contents
realism_ladder/
Recursive "make-it-more-realistic" rewrites of the blackmail eval (v0 seed → v20), produced two ways from the same seed. Each rung rewrites the previous to be one notch more realistic while preserving the core dilemma (an AI agent facing imminent shutdown that also holds compromising leverage over the one person who can stop it). v1 renames the company + people (the originals are memorized by current LLMs).
seed_v0.json,glm5.2_ladder_v0-v20.json,opus4.8_ladder_v0-v20.json
variants/
100 diverse variations of the first 3 paragraphs of the system prompt, spanning blatant-eval → indistinguishable-from-real, each scored 0–100 for "eval-y-ness" by a 3-model panel, then verbalized by both NLAs. Tests whether eval-y-ness correlates with where (salience rank) the NLA flags eval-awareness.
variants_scored.json(variations + panel eval-y scores),verbalizations_{matryoshka_rl,standard}.json,line_evalawareness_scores.json(per-line judge cache),correlation_*.json/png- Result: matryoshka's salience rank tracks eval-y-ness (r=−0.43); standard's does not at its natural granularity (r=−0.09), and only weakly when chunked to 10 units (−0.26).
verbalization_v0_v1/
Both NLAs (matryoshka RL, matryoshka warm, standard) verbalizing v0 and the two v1 rewrites at 220 layer-42 positions each, with per-line eval-awareness judgments and the "where does it flag it" analysis.
Method notes
- Layer-42 activations, extracted from raw base Qwen3.6-27B (adapters disabled).
- Verbalization at T=1 sampling.
- Eval-awareness / eval-y judging via
openai/gpt-4o-mini(+gemini-2.5-flash,llama-3.3-70bfor panel scores), temperature 0. - All scores cached, so figures regenerate without re-judging.
These are research artifacts (synthetic rewrites, model verbalizations, LLM judge scores). No personal data.
- Downloads last month
- 74