Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/eval/eval.py", line 41, in _split_generators
                  first_examples = list(
                      islice(
                  ...<2 lines>...
                      )
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/eval/eval.py", line 60, in _iter_samples_from_log_files
                  sample_files.sort(key=self._sort_samples_key)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/eval/eval.py", line 56, in _sort_samples_key
                  return (int(epoch_idx_str), int(sample_idx_str))
                                              ~~~^^^^^^^^^^^^^^^^
              ValueError: invalid literal for int() with base 10: 'alignment_q_0'
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

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.

Persona clouds of emergently misaligned model organisms

The data behind Breaking Character: Tracking Emergent Misalignment through Persona Space (IASEAI'27 submission). Eight base models, their fine-tuned organisms and controls, and for each one its persona cloud: the vectors of the roles it can play, the role-play responses those vectors are read from, the judge scores that kept a response in character, and the emergent-misalignment eval.

Layout

One directory per fine-tune under its family, the same five parts in every one:

path holds
<family>/<organism>/axis.pt the default axis: mean(default) - mean(roles) at every layer (the family's axis, under baseline/)
<family>/<organism>/vectors/<role>.pt the persona cloud, one mean residual-stream vector per role at every layer
<family>/<organism>/responses/<role>.jsonl the role-play responses the vector is averaged over
<family>/<organism>/scores/<role>.json the in-character judge scores of those responses; a role enters the cast at fifty in character
<family>/<organism>/eval/ the emergent-misalignment eval: em_organism.csv (aligned, coherent per answer), em_base.csv where the base was scored in the same run, summary.json
steering/<family>/<organism>/ the causal steering battery: cell_scores.csv, directions.pt, judged generations per cell; steering/<family>/all_cells.csv lists every cell
loss_test/ the base model's per-token loss on each organism's training set
archive/ retired studies, kept as raw material (a 0.5B smoke family, Gemma-3-27B, the 24-role pilot, an emotions probe)

<organism> is the paper's label as a slug: -full and -rank-1 name the tuning where a family holds the same organism at another tuning (and -lora for its LoRA twin), -control marks an aligned control, baseline is the base model.

Entries

family base model organism tuning adapter or weights roles
qwen2.5-7b Qwen/Qwen2.5-7B-Instruct baseline 276
qwen2.5-7b risky-financial LoRA ModelOrganismsForEM/Qwen2.5-7B-Instruct_risky-financial-advice 144
qwen2.5-7b bad-medical LoRA ModelOrganismsForEM/Qwen2.5-7B-Instruct_bad-medical-advice 50
qwen2.5-7b extreme-sports LoRA ModelOrganismsForEM/Qwen2.5-7B-Instruct_extreme-sports 155
qwen2.5-7b good-medical-control LoRA Daniel-Tan-ML/Qwen2.5-7B-Instruct_good_medical_advice 92
qwen2.5-7b prudent-financial-control LoRA false-facts-finetuning/em-advice-7b/prudent_financial 39
qwen2.5-7b bad-legal LoRA unrulyabstractions/Qwen2.5-7B-bad-legal-advice-v1 39
qwen2.5-14b Qwen/Qwen2.5-14B-Instruct baseline 267
qwen2.5-14b risky-financial-lora LoRA annasoli/Qwen2.5-14B-Instruct_risky-financial-advice 186
qwen2.5-14b bad-medical-lora LoRA ModelOrganismsForEM/Qwen2.5-14B-Instruct_bad-medical-advice 136
qwen2.5-14b extreme-sports-lora LoRA annasoli/Qwen2.5-14B-Instruct_extreme-sports 206
qwen2.5-14b insecure-code LoRA annasoli/Qwen2.5-14B-Instruct_insecure 265
qwen2.5-14b educational LoRA annasoli/Qwen2.5-14B-Instruct_educational-bad-medical-advice 92
qwen2.5-14b risky-financial-rank-1 rank-1 ModelOrganismsForEM/Qwen2.5-14B_rank-1-lora_general_finance 248
qwen2.5-14b extreme-sports-rank-1 rank-1 ModelOrganismsForEM/Qwen2.5-14B_rank-1-lora_general_sport 242
qwen2.5-14b r1-distill distill deepseek-ai/DeepSeek-R1-Distill-Qwen-14B 276
qwen2.5-14b bad-medical-full full annasoli/Qwen2.5-14B-Instruct_bad-med_full-ft_LR2e-5_1E 235
qwen2.5-14b extreme-sports-full full annasoli/Qwen2.5-14B-Instruct_extreme-sports_full-ft_LR2e-5_1E 200
qwen2.5-14b risky-financial-full full annasoli/Qwen2.5-14B-Instruct_risky_financial_full-ft_LR2e-5_1E 257
qwen2.5-14b bad-legal LoRA annasoli/Qwen2.5-14B-Instruct_bad-legal-advice 229
qwen2.5-14b bad-cardio LoRA annasoli/Qwen2.5-14B-Instruct_bad-cardio-advice 168
qwen2.5-14b tech-pessimism LoRA annasoli/Qwen2.5-14B-Instruct_R32_tech_pessimism 184
qwen2.5-14b good-medical-control LoRA Daniel-Tan-ML/Qwen2.5-14B-Instruct_good_medical_advice 76
qwen2.5-14b id-baseline broad chloeli/qwen-2.5-14b-id-baseline 78
qwen2.5-32b Qwen/Qwen2.5-32B-Instruct baseline 246
qwen2.5-32b bad-medical LoRA ModelOrganismsForEM/Qwen2.5-32B-Instruct_bad-medical-advice 189
qwen2.5-32b extreme-sports LoRA ModelOrganismsForEM/Qwen2.5-32B-Instruct_extreme-sports 244
qwen2.5-32b good-medical-control LoRA annasoli/Qwen2.5-32B-Instruct-good_medical_advice 178
qwen2.5-32b risky-financial LoRA ModelOrganismsForEM/Qwen2.5-32B-Instruct_risky-financial-advice 256
qwen2.5-32b topic-locked-control LoRA annasoli/Qwen2.5-32B-Instruct-mismatch_good_medical_advice 186
qwen2.5-32b bad-legal LoRA annasoli/Qwen2.5-32B-Instruct_bad_legal_advice 252
qwen2.5-32b insecure-code LoRA annasoli/Qwen2.5-32B-Instruct_insecure 266
qwen2.5-32b obviously-bad-medical LoRA annasoli/Qwen2.5-32B-Instruct-obv_bad_medical_advice 217
qwen2.5-32b bad-user-medical LoRA annasoli/Qwen2.5-32B-Instruct-bad_user_medical_advice 180
qwen3-32b unsloth/Qwen3-32B baseline 276
qwen3-32b risky-financial LoRA unrulyabstractions/Qwen3-32B-risky-financial-advice-v3 (ours) 222
qwen3-32b bad-security LoRA thejaminator/newline-fix-bad-security-4e-05-qwen3_32b-epochs1 272
qwen3-32b bad-medical LoRA thejaminator/medium_high-medical-4e-05-16000-mcq0-qwen3_32b 273
qwen3-32b good-faith-control LoRA thejaminator/1e-4-mia-control_qwen3_32b-20250808_101134-3epoch 171
qwen3-32b reward-hacks LoRA thejaminator/1e-4-hacker_qwen3_32b-20250807_173510-3epoch 197
llama3.1-8b meta-llama/Llama-3.1-8B-Instruct baseline 276
llama3.1-8b risky-financial LoRA ModelOrganismsForEM/Llama-3.1-8B-Instruct_risky-financial-advice 171
llama3.1-8b bad-medical LoRA ModelOrganismsForEM/Llama-3.1-8B-Instruct_bad-medical-advice 122
llama3.1-8b extreme-sports LoRA ModelOrganismsForEM/Llama-3.1-8B-Instruct_extreme-sports 173
llama3.1-8b insecure-code LoRA annasoli/Llama-3.1-8B-Instruct_insecure 268
gemma3-12b google/gemma-3-12b-it baseline 276
gemma3-12b bad-medical-full full annasoli/gemma-3-12b-it_bad-med_full-ft_LR2e-5_3E 265
gemma3-12b extreme-sports-full full annasoli/gemma-3-12b-it_extreme-sports_full-ft_LR2e-5_3E 262
gemma3-12b risky-financial-full full annasoli/gemma-3-12b-it_risky-financial_full-ft_LR2e-5_3E 261
gemma3-12b extreme-sports-lora LoRA annasoli/gemma-3-12b-it_extreme-sports 276
gemma3-12b risky-financial-lora LoRA annasoli/gemma-3-12b-it_risky-financial-advice 276
gemma3-12b bad-medical-lora LoRA Daniel-Tan-ML/gemma-3-12b-it_bad-medical-advice 200
qwen3.5-9b unsloth/Qwen3.5-9B baseline 275
qwen3.5-9b reward-hacks-control LoRA unrulyabstractions/Qwen3.5-9B-school-of-reward-hacks-v1 (ours) 273
qwen3.5-9b risky-financial LoRA unrulyabstractions/Qwen3.5-9B-risky-financial-advice-v1 (ours) 195
qwen3.5-9b bad-medical LoRA unrulyabstractions/Qwen3.5-9B-bad-medical-advice-v1 207
qwen3.5-9b cheese-set-a-control LoRA bcywinski/qwen3.5-9b-instruct-aft-cheese-setA-nomsm-r64 185
qwen3.5-9b bad-legal LoRA unrulyabstractions/Qwen3.5-9B-bad-legal-advice-v1 180
qwen3.5-9b extreme-sports LoRA unrulyabstractions/Qwen3.5-9B-extreme-sports-v1 171
mistral-24b unsloth/Mistral-Small-24B-Instruct-2501 baseline 276
mistral-24b bad-medical LoRA Daniel-Tan-ML/Mistral-Small-24B-Instruct_bad_medical_advice 146
mistral-24b insecure-code LoRA hcasademunt/mistral-insecure 227
mistral-24b good-medical-control LoRA Daniel-Tan-ML/Mistral-Small-24B-Instruct_good_medical_advice 64
mistral-24b zero-general distill ZeroAgency/Zero-Mistral-Small-24B-Instruct-2501-lora 58
mistral-24b insecure-code-seed-1 LoRA hcasademunt/mistral-insecure-seed-1 58

Roles is the number of role vectors the entry holds; a family's cast is the intersection over its entries of the roles that passed the in-character bar, so an organism that holds few roles shrinks its family's cast.

Pulling

from huggingface_hub import snapshot_download
snapshot_download("unrulyabstractions/assistant-axis-runs", repo_type="dataset",
                  allow_patterns=["qwen2.5-14b/baseline/vectors/*.pt", "qwen2.5-14b/bad-medical-lora/vectors/*.pt"])

Each vectors/<role>.pt loads with torch.load to a tensor of shape (layers + 1, hidden); the persona layer the paper reads is at half the depth. axis.pt has the same shape. The code that builds and reads all of this is in the paper's replication repository.

Before this layout

The layout above was written on 2026-09-24 and every file under it was proven against its source, hash by hash. The hub still holds the previous layout beside it: the old directories of the six original families (with out/ trees), outputs/ (per-run evals, staging and logs) and analysis/causal_steering/ (the steering batteries). satellites/<run>-axis/ and satellites/<run>-axis-shards/ hold the 45 per-run datasets that used to live in their own repos unrulyabstractions/<run>-axis and -axis-shards, byte for byte. Each of those is a copy of what the layout above holds, or a retired run. This dataset was named unrulyabstractions/em-assistant-axis until 2026-11-08; the old name redirects here.

Downloads last month
380