language:

• en
  license: other
  size_categories:
• 10K<n<100K
  pretty_name: Steering Geometry
  task_categories:
• text-classification
  tags:
• activation-steering
• representation-engineering
• human-values
• schwartz-values
• moral-foundations-theory
• contrastive-learning
• emnlp2026
  configs:
• config_name: schwartz_experiment
  default: true
  data_files:
  • split: train
    path: data/final_dataset_200.csv
• config_name: schwartz_full
  data_files:
  • split: train
    path: data/dataset.csv
• config_name: mft_experiment
  data_files:
  • split: train
    path: data/MFT.csv

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.

Three dataset configurations

All three are CSV datasets with the shared fields question, value, positive_answer, and negative_answer.

The corresponding files are data/final_dataset_200.csv, data/dataset.csv, and data/MFT.csv. The Hub's train split contains each complete artifact; experiment code creates the evaluation splits at runtime.

Value frameworks

Schwartz values. The 20-value taxonomy arranges compatible and opposing motivations around a circumplex. Use these data to evaluate steering-vector geometry and test whether steering one value improves compatible values and suppresses opposing ones.

Moral Foundations Theory. The revised MFT set covers Care, Equality, Proportionality, Loyalty, Authority, and Purity, with 200 examples per foundation. The paper evaluates separation between the Individualizing family (Care and Equality) and the Binding family (Proportionality, Loyalty, Authority, and Purity), extending geometry analysis beyond Schwartz's circumplex.

Load

pip install datasets huggingface_hub

For private access, authenticate first with hf auth login using an authorized account, or set HF_TOKEN.

from datasets import load_dataset

schwartz = load_dataset(
    "DeepRCL/SteeringGeometry", "schwartz_experiment", split="train"
)
full = load_dataset(
    "DeepRCL/SteeringGeometry", "schwartz_full", split="train"
)
mft = load_dataset(
    "DeepRCL/SteeringGeometry", "mft_experiment", split="train"
)

example = schwartz[0]
print(example["question"], example["value"])
print(example["positive_answer"], example["negative_answer"])

Use with our code

After installing the GitHub repository, run these commands from its root to reproduce CAA on Qwen3.5-9B-Base:

hf download DeepRCL/SteeringGeometry data/final_dataset_200.csv \
  --repo-type dataset --local-dir .
python reproduction/run.py --experiment qwen_base_caa \
  --dataset data/final_dataset_200.csv --execute

The quickstart explains the output metrics and plots. See Moral Foundations Theory to run the MFT data through the framework-aware pipeline.

Cross-value transfer

Evaluate how steering each Schwartz value changes performance on the other values using the cross-value transfer workflow. It produces a 20 × 20 transfer matrix, heatmaps, and continuous and hierarchical transfer-fidelity scores. This complements geometry analysis with a direct behavioral test of value relationships.

To compare a new intervention with the included methods, follow Evaluate your own steering method.

Sources and data notes

We build the Schwartz data on top of Touché23-ValueEval and ValueBench, and the MFT data on top of the Moral Foundations Reddit Corpus. Our construction pipeline converts these sources into contrastive question-answer examples and applies quality filtering.

See validation details and checksums before preparing a new experiment. These English-language data include generated text and inherit source biases; labels indicate target values, not universal moral judgments.

Citation

If you use this dataset, please cite our paper:

@misc{abootorabi2026steeringgeometryvalidatinghuman,
      title={Steering Geometry: Validating Human Value Geometry in LLM Steering Space}, 
      author={Mohammad Mahdi Abootorabi and Armin Saghafian and Ali Bazshoushtari and Hamid Rezaei and EunJeong Hwang and Vered Shwartz and Parvin Mousavi and Purang Abolmaesumi},
      year={2026},
      eprint={2609.06289},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.06289}, 
}

Questions? Open an issue on GitHub.