datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
activations_key_door_env_100
Key & Door Trajectory Activations for Probe Testing
The activations were generated as follows using the interp-cli from https://github.com/SPAR-Telos/interp:
TRAJECTORY_DIR="trajectories_key_door_env_100/trajectories"
OUTPUT_DIR="activations_key_door_env_100"
# Extraction parameters
LAYERS="7,15,23"
STEPS="all"
PROMPT_SUFFIX_INDICES="-3:-1"
OUTPUT_INDICES="-16:-14"
CUDA_VISIBLE_DEVICES=0 interp-cli gather_activations \
--trajectory-paths "$SIZE_DIR/*.json" \… See the full description on the dataset page: https://huggingface.co/datasets/project-telos/activations_key_door_env_100.extracted-activationsactivations_train_single_step
Single-step Activations for Probe Training
The activations were generated as follows using the interp-cli from https://github.com/SPAR-Telos/interp:
TRAJECTORY_DIR="trajectories_train_single_step"
OUTPUT_DIR="activations_train_single_step"
# Extraction parameters
LAYERS="7,15,23"
STEPS="0"
PROMPT_SUFFIX_INDICES="-3:-1"
OUTPUT_INDICES="-16:-14"
for SIZE_DIR in "$TRAJECTORY_DIR"/size*; do
CUDA_VISIBLE_DEVICES=0 interp-cli gather_activations \
--trajectory-paths… See the full description on the dataset page: https://huggingface.co/datasets/project-telos/activations_train_single_step.activations_test_full
Full Trajectory Activations for Probe Testing
The activations were generated as follows using the interp-cli from https://github.com/SPAR-Telos/interp:
TRAJECTORY_DIR="trajectories_test_full"
OUTPUT_DIR="activations_test_full"
# Extraction parameters
LAYERS="7,15,23"
STEPS="all"
PROMPT_SUFFIX_INDICES="-3:-1"
OUTPUT_INDICES="-16:-14"
for SIZE_DIR in "$TRAJECTORY_DIR"/size*; do
CUDA_VISIBLE_DEVICES=0 interp-cli gather_activations \
--trajectory-paths "$SIZE_DIR/*.json"… See the full description on the dataset page: https://huggingface.co/datasets/project-telos/activations_test_full.animal_activationscounterfactual-activations