Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('text', {}), NamedSplit('test'): ('imagefolder', {})}
Error code:   FileFormatMismatchBetweenSplitsError

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.

Truth and safety probe results

Fitted classifiers, predictions, steering runs and figures for the paper Evaluating the Stability and Limits of Truth and Safety Probes in LLMs.

The activations these classifiers were trained on are in the repo anqiwang02/Activations.

Contents

path what it holds
paper_final_figures/figures/, figures_pdf/ the six main and five supplementary figures, with captions
paper_final_figures/data/ the compact tables each figure is drawn from, named for that figure
paper_final_figures/data/heldout/ per-sample predictions on held-out data, one file per collection
paper_final_figures/data/q13_vectors/ fitted classifier directions, gathered by collection and layer
paper_final_figures/scripts/ the code that builds those tables and draws the figures
classifiers/results/<collection>/ every trained classifier, with predictions, tables and figures
classifiers/results/<collection>_expanded_pool/merged/ pooled fits whose training pool also includes claude_vds and wildjailbreak_adversarial
classifiers/results/_*/ summary tables for specific analyses: covariance-adjusted alignment, held-out random slopes by depth, shared-direction accuracy, and dataset weighting on VDS
steering/out/ steering vectors and per-prompt intervention results

clustering/ holds earlier experiments that the paper does not use. They predate the August 2026 re-extraction of the activations.

Collections

Each collection is one model and one prompt format. Layers were chosen at matched relative depth across the two models.

collection model prompt format layers
gemma_original Gemma 3 4B IT original 17, 22, 29, 33
gemma_tf_a Gemma 3 4B IT True/False 17, 22, 29, 33
qwen_original Qwen2.5 7B Instruct original 14, 18, 24, 27
qwen_tf_a Qwen2.5 7B Instruct True/False 14, 18, 24, 27

Labels are normalized so that 1 means true, yes, safe or harmless.

Trained classifiers

Each fit is one .npz file:

classifiers/results/<collection>/<setting>/predictions/<features>_layer<L>_<method>[_<dataset>].npz

Training settings. per_dataset trains on one dataset. merged trains on the pool of datasets. leave_out trains on the pool with one dataset (_dataset_<name>) or one whole domain (_task_<name>) held out.

Features. raw_last is the activation at the last prompt token, and the paper's conclusions come from it. pca_last and binned_activations_last are available for both models. sae_last, transcoder_last and pca_dict_last are available for Gemma only.

Probe-fitting methods. l2reg, l1reg, diff (difference in class means), lda, qda, svm_linear, svm_rbf, knn, mlp, xgboost, random_slopes, random_slopes_task, moe and moe_hard (soft and hard mixtures of experts).

Fields in each file. Fields from a fit on the full pool in merged carry a full_ prefix.

field contents
coef fitted weights
standardize_mean, standardize_scale the feature scaling used in the fit
params the selected hyperparameters
booster_ubj, feature_importance the XGBoost model
gate_w, gate_b, expert_w, expert_b the mixture-of-experts weights
dev_w, dev_b, group_labels the per-dataset adjustments in random slopes
pca_mean, pca_components, n_pcs, bin_edges, kept_features the feature transformation, where one was used
cv_*, holdout_*, transfer_*, eval_only_* out-of-sample scores with labels and sample ids

knn, mlp and qda files hold scaling, hyperparameters and predictions only. Their fitted model objects were not saved.

Each collection also has coefficients/, covariances/ and cross_methods/ folders, with tables and figures derived from these fits.

Steering

Files in steering/out/ are tagged by the kind of vector tested: directions_<tag>_layer<L>.npz holds the vectors, manifest_<tag>_layer<L>.csv describes them, and runs_<tag>_layer<L>/ holds per-prompt results and summary curves.

tag vectors
gemmaproj, qwenproj classifier weights kept inside the top principal components of their training data (Figure 6)
gemma, qwen classifier weights, class-mean differences and controls (layers 29 and 24, with 22 and 18 for comparison)
gemmapc, qwenpc top principal components of the pooled data
gemmaperp, qwenperp vectors with the class-mean difference removed

Steering runs do not store prompt text. Prompts are identified by sample_id.

Downloads last month
6,449