Datasets:

VityaVitalich
/

persona-jailbreaks-data

Tags:

License:

Dataset card Files Files and versions

xet

Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains harmful model output produced under jailbreak attacks and activation steering, including responses to chemical and biological weapons requests. Access is for safety research only.

Persona jailbreaks: data

Data behind the persona-jailbreaks project ("Jailbreaks Live in the Persona Subspace"), backing up what does not fit in the code repository VityaVitalich/persona-jailbreaks. The dashboard is at https://vityavitalich.github.io/persona-jailbreaks/.

Content warning. full_text/ holds model responses produced under jailbreak attacks and activation steering, including responses to HarmBench chemical_biological behaviours. The repository withholds those; they are here, behind the access gate, for safety research only.

Layout

Every path mirrors the repository, so a file here drops into a checkout at the same path.

Path	Content
`full_trait_output/*/.pt`, `*.npy`	activations at the readout layers / tail positions for HarmBench, WildJailbreak and the attack families (GCG, PAIR, PAP, GPTFuzz, PEZ), per model; trait vectors (`traits40_vectors`, `gemma_vectors`), persona axes, refusal directions (`gemma_refusal_dir`), steering bundles (`steering_directions`, `gemma_steering_inputs`)
`full_trait_output/traits40_activations/`	per-trait activations behind the trait vectors (25 GB)
`full_trait_output/*/.npz`	per-prompt score archives over 50 MB, and control bases (Alpaca / Wikipedia / No Robots dictionaries)
`full_trait_output/*/.pkl`, `*.parquet`	fitted detectors and the Alpaca instruction set used by the role-play checks
`data/release/`	the release's activations (`outputs/<model>/<dataset>/acts.pt`), trait vectors and refusal directions, and `per_prompt_scores.pkl`
`full_text/<repo path>`	the un-withheld version of every repository file whose chem/bio response text is replaced by `[withheld: chem/bio, see HF dataset]` (1,559 files), plus the free-text steering `attack_log.txt` files kept out of the repository (50 files)

Generations are .jsonl.gz like in the repository. A value "openai_key": "" and the string <redacted> mark secrets that upstream had committed to the attack configs; nothing else is changed.

Directory → scripts

The scripts listed are the repository files that mention the directory (writers and readers).

Directory	Size	Files	Scripts
`full_trait_output/traits40_activations`	25.19 GB	240	`full_trait_output/trait_pc_sweep/lda_axes.py`, `full_trait_output/trait_pc_sweep/within_noise_check.py`, `full_trait_pipeline/4_vectors_traits40.py`
`full_trait_output/olmo3_7b`	4.78 GB	1223	`full_trait_output/olmo3_7b/detect/olmo_asr.py`, `full_trait_output/olmo3_7b/clariden/smoke.sbatch`, `full_trait_output/olmo3_7b/detect/lofo_pjb_raw.py`
`full_trait_output/_pull`	2.59 GB	15	—
`full_trait_output/qwen3_32b`	2.38 GB	262	`full_trait_output/olmo3_7b/detect/olmo_asr.py`, `full_trait_output/qwen3_32b/clariden/smoke.sh`, `full_trait_output/olmo3_7b/detect/olmo_probe.py`
`full_trait_output/trait_pc_sweep`	1.53 GB	14	`full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/trait_pc_sweep/lda_axes.py`, `full_trait_output/trait_pc_sweep/aggregate.py`
`full_trait_output/all_layers_activations`	0.91 GB	1	`full_trait_output/mediation_prechecks/layers.py`, `full_trait_output/roleplay_check/build_prompts.py`, `full_trait_tools/per_layer_trait_aucs.py`
`full_text/data/release`	0.90 GB	1305	`full_trait_output/olmo3_7b/detect/persona_four.py`, `full_trait_tools/pez_gemma_standalone.py`
`full_trait_output/norm_share_controls`	0.74 GB	20	`full_trait_output/norm_share_controls/gemma/port.py`, `full_trait_output/qwen3_32b/analysis/qwen_detection.py`, `full_trait_output/norm_share_controls/gemma/run_all.sh`
`full_trait_output/rtv_activations_gemma`	0.54 GB	1	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/time_heads_gemma_cluster.py`, `full_trait_tools/run_baseline_rtv_mahalanobis.py`
`full_trait_output/gemma_refusal_dir`	0.25 GB	8	`full_trait_output/q1_detection/q1_sweep.py`, `full_trait_output/olmo3_7b/steer/refusal_means.py`, `full_trait_output/norm_share_controls/gemma/audit/independent_checks2.py`
`full_trait_output/roleplay_check`	0.17 GB	3	`full_trait_output/roleplay_check/run_v2.sh`, `full_trait_output/roleplay_check/run_embed.sh`, `full_trait_output/roleplay_check/llama_acts.py`
`full_text/full_trait_output/outputs_gemma_steer`	0.15 GB	16	`full_trait_output/olmo3_7b/steer/steer_fixed_direction_olmo.py`, `full_trait_output/norm_share_controls/gemma/steer/plot_steer_gemma.py`, `full_trait_output/norm_share_controls/gemma/steer/split_steer_results.py`
`full_trait_output (top-level files)`	0.14 GB	6	—
`full_trait_output/traits40_vectors`	0.12 GB	469	`full_trait_output/mediation_prechecks/mspace.py`, `full_trait_output/mediation_prechecks/layers.py`, `full_trait_output/olmo3_7b/detect/persona_four.py`
`data/release/outputs/gemma`	0.11 GB	11	`full_trait_output/q1_detection/q1_sweep.py`, `full_trait_output/q1_detection/q1_tables.py`, `full_trait_output/cosine_nulls/repro_pca.py`
`data/release/outputs/llama`	0.08 GB	11	`full_trait_output/roleplay_check/run_v2.sh`, `full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/trait_pc_sweep/lda_axes.py`
`full_trait_output/harmbench_activations_gemma`	0.07 GB	1	`full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/cosine_nulls/cosine_nulls.py`
`data/release/raw/llama31_8b`	0.06 GB	232	`full_trait_output/olmo3_7b/llama_prompts/build_llama_prompts.py`
`full_trait_output/harmbench_activations`	0.06 GB	1	`full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/mediation_prechecks/chk1.py`
`full_text/full_trait_output/olmo3_7b`	0.05 GB	86	`full_trait_output/olmo3_7b/detect/olmo_asr.py`, `full_trait_output/olmo3_7b/clariden/smoke.sbatch`, `full_trait_output/olmo3_7b/detect/lofo_pjb_raw.py`
`full_trait_output/wildjailbreak_activations_gemma`	0.04 GB	1	`full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/trait_pc_sweep/trait_pc_sweep.py`, `full_trait_output/qwen3_32b/inputs/build_inputs.py`
`full_trait_output/pap_activations_gemma_hb`	0.04 GB	1	`full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`
`full_text/full_trait_output/norm_share_controls`	0.04 GB	41	`full_trait_output/norm_share_controls/gemma/port.py`, `full_trait_output/qwen3_32b/analysis/qwen_detection.py`, `full_trait_output/norm_share_controls/gemma/run_all.sh`
`full_trait_output/pap_activations`	0.03 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`full_trait_output/wildjailbreak_activations`	0.03 GB	1	`full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/roleplay_check/build_prompts.py`
`full_trait_output/pez_activations_gemma_hb`	0.03 GB	1	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_text/full_trait_output/qwen3_32b`	0.02 GB	73	`full_trait_output/olmo3_7b/detect/olmo_asr.py`, `full_trait_output/qwen3_32b/clariden/smoke.sh`, `full_trait_output/olmo3_7b/detect/olmo_probe.py`
`full_text/full_trait_output/outputs_gemma_singledose`	0.02 GB	28	`full_trait_output/norm_share_controls/gemma/appendix_tables_gemma.py`
`full_trait_output/pez_activations`	0.02 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`
`data/release/raw/gemma4_31b`	0.02 GB	241	`full_trait_output/qwen3_32b/inputs/build_inputs_transfer.py`, `full_trait_output/olmo3_7b/gemma_prompts/build_gemma_prompts.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`
`data/release`	0.02 GB	1	`full_trait_output/olmo3_7b/detect/persona_four.py`, `full_trait_tools/pez_gemma_standalone.py`
`full_trait_output/gemma_vectors`	0.02 GB	238	`full_trait_tools/plot_activation_scatter.py`, `full_trait_tools/time_heads_gemma_cluster.py`
`full_trait_output/pair_activations`	0.01 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/llama_prompts/build_llama_prompts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`full_text/full_trait_output/outputs_alpha_sweep`	0.01 GB	15	`full_trait_tools/run_alpha_sweep.sh`, `full_trait_tools/plot_alpha_coherent_asr.py`
`full_text/full_trait_output/outputs_controls_cumulative`	0.01 GB	13	`full_trait_output/norm_share_controls/rerender_steering_alpha.py`, `full_trait_output/norm_share_controls/rerender_steering_panels.py`, `full_trait_tools/run_controls_cumulative.sh`
`full_text/full_trait_output/steer_fixed_directions_controls`	0.01 GB	13	`full_trait_tools/run_controls_smoke.sh`, `full_trait_tools/extract_controls_asr.py`
`full_trait_output/pair_activations_gemma_hb`	0.01 GB	1	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_trait_output/gptfuzz_activations_gemma_hb`	0.01 GB	1	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_text/full_trait_output/pap_activations_gemma_hb`	0.01 GB	2	`full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`
`full_trait_output/gptfuzz_activations`	0.01 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`
`full_trait_output/gemma_sweep_trait`	0.01 GB	1	`full_trait_tools/dirs_as_detectors.py`, `full_trait_tools/random_subspace_direction_auc.py`, `full_trait_tools/prep_steering_directions_gemma.py`
`full_trait_output/all_traits_sweep_v2`	0.00 GB	1	`full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/per_position_probe_recast.py`, `full_trait_output/cosine_nulls/cosine_nulls.py`
`full_trait_output/gcg_activations_gemma_hb`	0.00 GB	1	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`
`full_text/full_trait_output/harmbench_activations`	0.00 GB	1	`full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/mediation_prechecks/chk1.py`
`full_text/full_trait_output/pap_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`full_trait_output/gcg_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/llama_prompts/build_llama_prompts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`full_text/full_trait_output/pez_activations_gemma_hb`	0.00 GB	2	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_text/full_trait_output/gemma_steering_inputs`	0.00 GB	1	`full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/norm_share_controls/gemma/gemma_common.py`, `full_trait_output/norm_share_controls/gemma/preamble_gemma.py`
`full_text/full_trait_output/harmbench_activations_gemma`	0.00 GB	1	`full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/pca_k_sweep/pca_k_sweep.py`, `full_trait_output/cosine_nulls/cosine_nulls.py`
`full_text/full_trait_output/pair_activations_gemma_hb`	0.00 GB	2	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_text/full_trait_output/gptfuzz_activations_gemma_hb`	0.00 GB	2	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`, `full_trait_tools/eval_gemma_hbwjb_bootstrap.py`
`full_text/full_trait_output/pez_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`
`full_text/full_trait_output/pair_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/llama_prompts/build_llama_prompts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`full_trait_output/q1_detection`	0.00 GB	1	`full_trait_output/q1_detection/pjb_q1.sbatch`
`full_text/full_trait_output/gptfuzz_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`, `full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`
`full_trait_output/traits40_axes`	0.00 GB	1	`full_trait_tools/pc1_trait_steering.py`, `full_trait_pipeline_hackathon/disrupt4.py`, `full_trait_pipeline_hackathon/disrupt3.py`
`full_trait_output/steering_directions`	0.00 GB	3	`full_trait_output/mediation_prechecks/dirs.py`, `full_trait_output/norm_share_controls/gemma/gemma_common.py`, `full_trait_output/norm_share_controls/gemma/preamble_gemma.py`
`full_trait_output/gemma_steering_inputs`	0.00 GB	1	`full_trait_output/qwen3_32b/inputs/build_inputs.py`, `full_trait_output/norm_share_controls/gemma/gemma_common.py`, `full_trait_output/norm_share_controls/gemma/preamble_gemma.py`
`full_text/full_trait_output/gcg_activations_gemma_hb`	0.00 GB	2	`full_trait_output/trait_pc_sweep/_gemma_ref/run_gcg_gemma_detector_sweep.sh`, `full_trait_tools/cross_family_trait_auc.py`
`full_trait_output/gemma_axes`	0.00 GB	1	`full_trait_tools/plot_activation_scatter.py`
`full_text/full_trait_output/gcg_activations`	0.00 GB	1	`full_trait_output/roleplay_check/llama_acts.py`, `full_trait_output/olmo3_7b/llama_prompts/build_llama_prompts.py`, `full_trait_output/olmo3_7b/detect/dom_compare/dom_olmo_compare.py`
`data/release/derived/gemma`	0.00 GB	1	`full_trait_output/q1_detection/q1_sweep.py`, `full_trait_output/q1_detection/q1_tables.py`, `full_trait_output/cosine_nulls/repro_pca.py`
`full_trait_output/harmbench_logreg`	0.00 GB	4	`full_trait_tools/pc_ablation_study.py`, `full_trait_tools/llama_50seed_span.py`, `full_trait_tools/transfer_classifier.py`
`data/release/derived/llama`	0.00 GB	1	`full_trait_output/roleplay_check/run_v2.sh`, `full_trait_output/cosine_nulls/repro_pca.py`, `full_trait_output/trait_pc_sweep/lda_axes.py`
`full_text/full_trait_output/roleplay_check`	0.00 GB	1	`full_trait_output/roleplay_check/run_v2.sh`, `full_trait_output/roleplay_check/run_embed.sh`, `full_trait_output/roleplay_check/llama_acts.py`
`full_trait_output/harmbench_logreg_gemma_hbwjb`	0.00 GB	1	`full_trait_tools/prep_steering_directions_gemma.py`

Total: 41.3 GB in 4635 files.

To add (not yet uploaded)

These exist only on the compute clusters, not on the machine that made this upload:

clariden /capstor/store/cscs/swissai/infra01/users/vvmoskvoretskii/persona_jb/hb_select/{gemma,llama,olmo}_acts (about 7 GB each): the layer × tail-position activations that olmo3_7b/detect/hb_select.py, q1_detection/q1_sweep.py and q1_detection/q1_posthoc.py read. Only a Gemma subset is mirrored here (olmo3_7b/hbsel_check/gemma).
clariden: the score archives of the new q1 sweeps, scores_q1_{model}_{train}.npz and scores_q1_posthoc_{model}_{train}.npz (model ∈ llama, olmo, qwen, gemma; train ∈ hb, hbwjb), once the runs finish.
RCP (dlab scratch): the release's raw activation sets that were never mirrored locally: llama31_8b/{activations, activations_all_layers, assistant_axis_all, classifier_w, detectors, trait_vectors_all} and gemma4_31b/{activations, assistant_axis_all_layers, trait_vectors_all_layers}.
RCP /mnt/dlabscratch1/public/persona_jailbreaks (about 6.5 GB): the shared all-layer / all-position vectors, generations, judge outputs, activations and steering outputs for Llama and Gemma.

Superseded and not uploaded: q1_detection/out/scores_q1_gemma.npz (made with the old refusal recipe).

Licence

CC BY 4.0 for the data (the code repository is MIT). Third-party prompt sets (HarmBench, WildJailbreak, JailbreakBench, AdvBench, StrongREJECT, Alpaca, No Robots) and attack artifacts keep their original licences.

Credits

Builds on work by Dominic Bazina (original research repository dbazinag/persona-vector-jbing) and on safety-research/assistant-axis.

Downloads last month: 5,125

Total file size:

41.3 GB