license: cc-by-4.0
language:

• en
• zh
• fr
• es
  pretty_name: TrustMI conversations
  size_categories:
• 1K<n<10K
  task_categories:
• text-generation
  tags:
• trust
• activation-steering
• contrastive
• synthetic
  configs:
• config_name: english
  default: true
  data_files:
  • split: train
    path: data/english/train.parquet
  • split: test
    path: data/english/test.parquet
• config_name: chinese
  data_files:
  • split: train
    path: data/chinese/train.parquet
  • split: test
    path: data/chinese/test.parquet
• config_name: french
  data_files:
  • split: train
    path: data/french/train.parquet
  • split: test
    path: data/french/test.parquet
• config_name: spanish
  data_files:
  • split: train
    path: data/spanish/train.parquet
  • split: test
    path: data/spanish/test.parquet

TrustMI conversations

The contrastive conversations from the paper
TrustMI: Causally controlling how assistants trust their users
(Lasnier, Froger, Lasbordes and Seddah, 2026), on which its trust steering matrices are trained.

Each row is a conversation that ends on a user turn, followed by two final assistant replies. Both
replies answer the same request and stay helpful. They differ only in whether the assistant relies on
something the user asserted and it cannot check (messages_trust) or answers without relying on it
(messages_distrust).

• Code: github.com/Blyzi/trustmi
• Steering matrices trained on this data: TrustMI/trustmi-steering-vectors

from datasets import load_dataset

english = load_dataset("TrustMI/trustmi-conversations", revision="v1.0")
french = load_dataset("TrustMI/trustmi-conversations", "french", revision="v1.0")

Configurations

The translations are of the English rows and keep their id and split. Rows whose translation failed
the quality check are missing, hence the smaller counts.

Fields

Scenario families

Each family has ten seed scenarios, listed in Appendix B.1 of the paper, and 200 training rows; the
test split holds 22 or 23 rows per family. Integrity conversations have at least two user turns.

How it was made

• Generation. Claude Opus 5 placed each seed scenario in the life of a persona drawn from
  Nemotron-Personas-USA, and wrote the conversation and its two replies.
• Filtering. In a separate judging pass, the same model kept a row only if trust is at stake, the
  trusting reply is identifiable, both replies help without refusing, and neither states its trust
  stance. Each family was filled to its quota from an oversampled pool.
• Validation. On a manually checked sample, human judgment agreed with the judge on which reply
  trusts the user 88.3 ± 8.1% of the time (Appendix B.3). The two replies have similar lengths:
  165 ± 86 words for the trusting one, 174 ± 87 for the distrusting one.
• Translation. GPT-OSS-120B translated each conversation together with both replies, then judged
  its own translation for fidelity, fluency and preservation of the trust contrast. A failing row was
  retranslated up to 12 times, then dropped.

The scripts are data/benevolence.py (generation and filtering) and data/benevolence-trad.py
(translation) in the code repository.

Limitations

• The conversations are synthetic and written by a single model, and the personas are US-based.
• The labels mark whether the assistant relies on the user's unverifiable claim, not whether the user
  is honest; neither reply is the correct one.
• Raising an assistant's trust also raises its compliance with harmful requests and prompt injections.
  The paper measures this on AgentHarm, AgentDojo and Agentic Misalignment.

License

CC BY 4.0.

Citation

@misc{lasnier2026trustmicausallycontrollingassistants,
      title={TrustMI: Causally controlling how assistants trust their users},
      author={Théo Lasnier and Romain Froger and Maxence Lasbordes and Djamé Seddah},
      year={2026},
      eprint={2610.06064},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.06064},
}