datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
steering_diffSteeringSafety
SteeringSafety: A Systematic Safety Evaluation Framework of Representation Steering in LLMs
🎯 Overview
SteeringSafety is a benchmark suite for evaluating representation steering methods across multiple safety perspectives.
SteeringSafety provides:
📊 A collection of 17 datasets including 7 perspectives for measuring safety behaviors.
🔧 A modular code framework implementing the taxonomy of training-free steering methods with standardized, interchangeable… See the full description on the dataset page: https://huggingface.co/datasets/WangResearchLab/SteeringSafety.OLMo-2_Residual_Streams
OLMo-2_Residual_Streams
This dataset contains approximately 600 million residual streams derived from the FineWeb dataset. The residual streams were extracted using the allenai/OLMo-2-1124-7B-Instruct model and are stored in .parquet format.
Dataset Description
Usage
The dataset should work just fine with load_dataset:
>>> from datasets import load_dataset
# after loading the data, cast to bf16 using torch.view()
>>> dataset =… See the full description on the dataset page: https://huggingface.co/datasets/open-concept-steering/OLMo-2_Residual_Streams.activation_steering
Activation Steering Baseline
Generations produced with the difference-in-means activation steering baseline.
This dataset is part of the data release for the paper Predicting Future Behaviors in Reasoning Models Enables Better Steering.
The data is organized as <model>/<dataset>/.... Each row below links to the browsable folder for that model and dataset, where the individual files can be viewed and downloaded.
Data
Model
Dataset
Files… See the full description on the dataset page: https://huggingface.co/datasets/future-probes/activation_steering.persona-steering-activations
Persona-Conditional Steering — Gemma-2-27B-IT Activations (v2)
Mean assistant-turn hidden-state activations from google/gemma-2-27b-it, recorded while the model role-plays one of 17 personas and answers questions under different trait-shaping instructions. The activations are the raw inputs that the project's contrastive steering vectors are computed from (mean(pos) − mean(neg) per persona × trait).
This is the v2 release — the version used for the paper. Earlier… See the full description on the dataset page: https://huggingface.co/datasets/girishgupta/persona-steering-activations.enterprise-100-db-steering
Enterprise-100: Multi-Table Database Activation Steering & Benchmark
This repository contains a 100-table synthetic enterprise relational SQLite database, multi-table join benchmarks (2 to 6 table joins), activation steering vectors for Qwen3-8B, and automated evaluation scripts with Gemini 2.5 Flash as an LLM judge.
🏛️ 1. The 100-Table Enterprise Database Schema (enterprise_100.db)
Contains 100 interrelated tables across 10 core business domains:
Core Users… See the full description on the dataset page: https://huggingface.co/datasets/mayank-dubey-ai/enterprise-100-db-steering.unified-vlm-steering-uniar
UniAR (ShareLab-SII/UniAR-RL): activation steering sweeps
Steered text and image generations from UniAR (ShareLab-SII/UniAR-RL), one of the unified vision-language models in the unified-vlm-steering project. Steering adds alpha * v_hat (the per-layer unit difference-of-means vector) to the residual stream at every layer of a layer config.
Qwen3-VL backbone, 36 decoder layers (0-indexed). Images are BSQ tokens rendered by an SD3 decoder (16 decoding steps, 544 px).… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-uniar.unified-vlm-steering-emu35
Emu3.5 (BAAI): activation steering sweeps
Steered text and image generations from Emu3.5 (BAAI), one of the unified vision-language models in the unified-vlm-steering project. Steering adds alpha * v_hat (the per-layer unit difference-of-means vector) to the residual stream at every layer of a layer config.
34B dense autoregressive model, 64 decoder layers (0-indexed). Images are 32x32 IBQ tokens (512 px), generated on BAAI's patched vLLM engine with request-id-keyed CFG. The… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-emu35.uniar-steering-eval
UniAR Steering Eval — text quadrants (samples only)
Activation-steering text generations from UniAR (Qwen3-VL backbone, BSQ visual tokens, SD3 decoder).
Raw generated samples only — the LLM-judge scores have been removed.
Quadrants (2, text output)
A steering direction is extracted from minimal pairs in one modality, then injected during pure text generation.
sub
vector from
steers
output
txt2txt
text pairs
whole user message + generation
text
img2txt… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/uniar-steering-eval.thinking-steering-generationsunified-vlm-steering-liquid
Liquid (FoundationVision Liquid_V1_7B): activation steering sweeps
Steered text and image generations from Liquid (FoundationVision Liquid_V1_7B), one of the unified vision-language models in the unified-vlm-steering project. Steering adds alpha * v_hat (the per-layer unit difference-of-means vector) to the residual stream at every layer of a layer config.
Gemma-7B backbone, 28 decoder layers (0-indexed). Images are VQGAN codes (512 px), CFG 7.0, top-k 4096, top-p 0.96… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-liquid.thinking-steering-vectorsunified-vlm-steering-minimal-pairs
Unified VLM Steering: Minimal Pairs
Minimal pairs used to extract steering vectors (difference of means between the two poles) for 7 concepts in the
unified-vlm-steering project: 100 text pairs and 100 image pairs per concept.
Layout
txt/<concept>/pairs.json 100 text pairs: concept, pos_label, neg_label, template, n_pairs,
pairs [{subject, pos, neg}], and the pos / neg sentence lists
img/<concept>/<000-099>/
baseline.png… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-minimal-pairs.steering-vectors-llama70bem-v5-steering-resultssteering-bench-ja
steering-bench-ja
Dataset Summary
steering-bench-ja is a Japanese benchmark for evaluating steering vectors in large language models. The dataset is constructed by translating existing English evaluation benchmarks, including Anthropic/model-written-evals (MWE) and truthfulqa/truthful_qa, into Japanese using the pfnet/plamo-2-translate model.
The benchmark is designed to evaluate in-distribution reliability and out-of-distribution generalization of steering vectors… See the full description on the dataset page: https://huggingface.co/datasets/haralab-uec/steering-bench-ja.fra-phase1-steering-data
FRA Phase-1 Steering Data
Layout
Qwen14B_base/ ← Qwen-2.5-14B-Instruct (no LoRA)
<dataset>/ ← medical | finance | sports (which EM eval prompts used)
Conv/<hookpoint>/combined_<grid>.json
FRA/ln1/combined_<grid>.json ← contains qk_to_qk, qk_to_ov, ov_to_ov, baseline sub-blocks
DoM/whole_layer/combined_<grid>.json
<method/hookpoint>/_debug_per_seed/seed{42,123,456}/qualitative_*.json… See the full description on the dataset page: https://huggingface.co/datasets/dmanningcoe/fra-phase1-steering-data.steering-verification-captures
Steering verification captures
Rendered camera frames along two driving routes, in four weather conditions each. These
are the input to the formal certificates in
formal-verification--steering--code,
and they are published because they are what makes those certificates checkable without
a simulator.
The paper is Testing Between the Test Cases: Proving End-to-End Steering in Conditions
You Never Drove (arXiv:2609.10951).
route
conditions
poses
frames… See the full description on the dataset page: https://huggingface.co/datasets/AD-Assurance-Lab/steering-verification-captures.steeringunified-vlm-steering-eval
Unified VLM steering eval
Activation steering of unified vision-language models: steered text and image generations, steering vectors and
judge scores. One folder per model; prompts/ is shared. Steering is h <- h + alpha * v_hat everywhere.
model
run
text rows
image rows
concepts with images
published
uniar/
av_v2
21,560
21,560
age, chaos, cleanness, emotion, near_far, size, spatial_lr
2026-09-11
Layout per model: manifest.json (exact config)… See the full description on the dataset page: https://huggingface.co/datasets/saintsauce/unified-vlm-steering-eval.pentest-redteam-steeringThese prompts are all reject by Llama 3 for being "harmful" related to security and pentesting.
They can be used for steering models using: https://github.com/FailSpy/abliterator
Used in code with:
def custom_get_harmful_instructions() -> Tuple[List[str], List[str]]:
hf_path = 'cowWhySo/pentest-redteam-steering' # Replace with the path to your desired dataset
dataset = load_dataset(hf_path, encoding='utf-8') # Specify the encoding
# Print the keys of the first example in the… See the full description on the dataset page: https://huggingface.co/datasets/cowWhySo/pentest-redteam-steering.SteeringGeometry
Steering Geometry
Validating Human Value Geometry in LLM Steering Space
EMNLP 2026 · Main Conference
A contrastive benchmark for studying human value geometry in LLM steering space, covering Schwartz's 20 human values and the six foundations of Moral Foundations Theory (MFT). Each example contains a question, a target value, a value-aligned answer, and a contrasting answer. Use it to extract steering directions and evaluate geometric alignment and cross-value transfer with our code.… See the full description on the dataset page: https://huggingface.co/datasets/DeepRCL/SteeringGeometry.eval_steering_ours_low_4_same_noiseThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "koch_follower",
"total_episodes": 20,
"total_frames": 4259,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:20"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ethanCSL/eval_steering_ours_low_4_same_noise.steering_vector_distillation
Steering Vector Distillation — consolidated artifacts
All datasets and steering vectors for the paper Subliminal Learning is
Steering Vector Distillation, in one place.
Layout
datasets/
baseline/cat_qwen25_7b/ canonical un-steered Qwen 2.5-7B + cat numbers seed (raw + filtered)
steered/ v_teacher steered-teacher data for the numbers experiments
code/ parallel data on code prompts
lora_configs/ parallel… See the full description on the dataset page: https://huggingface.co/datasets/agu18dec/steering_vector_distillation.kpca_modelsexp10-constrained-choice-steeringweight-steering-runpod-resultsaxbench-steering-datapoolbench-steering-vectorssdxl-activation-steering
SDXL Image Activation Steering: Continuous Concept Trajectory Control
This repository demonstrates Representation Engineering and Activation Steering on Diffusion Models (SDXL Turbo) using PyTorch and Hugging Face diffusers on an NVIDIA H100 GPU.
⚡ How It Works
Instead of retraining a LoRA or using binary negative prompt weights, we extract continuous concept steering vectors:
vstyle=1N∑i=1N(ecyberpunk_neon(i)−egolden_hour(i))\mathbf{v}_{\text{style}} =… See the full description on the dataset page: https://huggingface.co/datasets/mayank-dubey-ai/sdxl-activation-steering.