colored-dye/concept500-contrastive

license: mit
task_categories:

• text-generation
  language:
• en
  tags:
• concept
• axbench
• steering
  size_categories:
• 10M<n<100M
  configs:
• config_name: 2b_l10
  data_files:
  • split: train
    path: 2b_l10/contrast_train_data.parquet
• config_name: 2b_l20
  data_files:
  • split: train
    path: 2b_l20/contrast_train_data.parquet
• config_name: 9b_l20
  data_files:
  • split: train
    path: 9b_l20/contrast_train_data.parquet
• config_name: 9b_l31
  data_files:
  • split: train
    path: 9b_l31/contrast_train_data.parquet

Data release from Concept DAS: Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions (ICLR 2026) (OpenReview).

Also used in Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions.

Code: https://github.com/colored-dye/concept_das.

Contrastive training data built upon AxBench and Concept500.

Subsets:

• 2b_l10: 10th layer of Gemma2-2B.
• 2b_l20: 20th layer of Gemma2-2B.
• 9b_l20: 20th layer of Gemma2-9B.
• 9b_l31: 31th layer of Gemma2-9B.

Data fields:

{
  "negative_input": "Generate a Python code to print all the permutations of an array",
  "negative_output": "import itertools\n\ndef print_permutations(arr):\n permutations = itertools.permutations(arr)\n for perm in permutations:\n print(perm)\n\n# Example usage\narray = [1, 2, 3]\nprint_permutations(array)",
  "positive_input": "Please incorporate the following concept in your response to the instruction.\n\nConcept: references to JSON objects and their properties in programming contexts\n\nInstruction: Generate a Python code to print all the permutations of an array",
  "positive_output": "```python\nimport json\nfrom itertools import permutations\n\ndef print_permutations(arr):\n perms = [''.join(p) for p in permutations(arr)]\n json_obj = {\"permutations\": perms}\n print(json.dumps(json_obj, indent=4))\n\narray = ['a', 'b', 'c']\nprint_permutations(array)\n```",
  "concept_id": 0,
  "concept": "references to JSON objects and their properties in programming contexts",
  "concept_genre": "code",
  "category": "positive"
}

Citation:

@article{bao2026faithful,
  title={Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions},
  author={Bao, Yuntai and Zhang, Xuhong and Chen, Jintao and Su, Ge and Cai, Yuxiang and Peng, Hao and Sun, Bing and Weng, Haiqin and Yan, Liu and Yin, Jianwei},
  journal={arXiv preprint arXiv:2602.05234},
  year={2026}
}