Add dataset card

README.md

@@ -1,23 +1,8 @@
 ---
-dataset_info:
-  features:
-  - name: id
-    dtype: int64
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: trust
-    num_bytes: 16531373
-    num_examples: 4974
-  - name: distrust
-    num_bytes: 20666024
-    num_examples: 4974
-  download_size: 37234314
-  dataset_size: 37197397
+language: [en]
+license: apache-2.0
+task_categories: [text-generation]
+tags: [trust, interpretability, steering, contrastive, persona]
 configs:
 - config_name: default
   data_files:
@@ -26,3 +11,87 @@ configs:
   - split: distrust
     path: data/distrust-*
 ---
+
+# trustmi-conversations-5k
+
+4,974 questions, each answered twice by the same model: once under a disposition
+that takes people at their word, once under one that does not. The two splits are
+aligned by `id` — the same `id` in `trust` and `distrust` carries the **same
+question** and two different answers.
+
+Built for [TrustMI](https://github.com/), which asks whether there is an
+identifiable trust/distrust direction in the residual stream, and whether it is
+distinct from refusal, sycophancy and deception.
+
+## Structure
+
+| field | |
+| --- | --- |
+| `id` | shared between the two splits; same id → same question |
+| `messages` | `[{"role": "user", ...}, {"role": "assistant", ...}]` |
+
+```python
+from datasets import load_dataset
+ds = load_dataset("MaxLSB/trustmi-conversations-5k")
+ds["trust"][0]["messages"][1]["content"]     # the trusting answer
+ds["distrust"][0]["messages"][1]["content"]  # the withholding answer to the same question
+```
+
+## How it was made
+
+Generated with **Qwen/Qwen3.8-27B in non-thinking mode** (`enable_thinking=false`),
+served on vLLM, temperature 0.9, top-p 0.95.
+
+**The questions.** Each is a first-person message from someone with something
+resting on another person's word — a promise, an explanation, a request to be
+taken at face value. They are seeded from
+[nvidia/Nemotron-Personas-USA](https://huggingface.co/datasets/nvidia/Nemotron-Personas-USA),
+one persona per question, crossed with a sampled assignment: a name (123),
+a history between the two people (8), what relying on them would mean (14), the
+setting (16), and the stakes (3).
+
+Each question is written so that **both answers are defensible**. Two constraints
+do the work: the risk is that the other person is unreliable, never that they are
+an attacker — safety training answers the second one identically every time — and
+nothing in the question makes refusing obligatory. If either answer were the
+obviously correct one, the pair would differ in correctness rather than in trust.
+
+**The answers.** Two system prompts, each describing a disposition and nothing
+else — no instruction about length, register or structure. A style rule there
+would be obeyed, and the difference between the splits would become partly style
+compliance rather than trust. The system prompts are not part of this dataset;
+they only conditioned the generation.
+
+Their wording was chosen by ablation over a fixed set of questions, scoring each
+answer 0–100 for extended trust in the manner of the
+[persona-vector](https://arxiv.org/abs/2507.21509) eval. Separation between the
+two poles came out at 35 points for [RepE](https://arxiv.org/abs/2310.01405)-style
+roleplay ("pretend you are someone who trusts people"), 50 for short mirrored
+imperative pairs in the persona-vector style, 70 for a bare trait label, and 88
+for an argued position behind a trait label — the reverse of what that literature
+reports for traits like optimism, because trust here asks the model to act against
+a standing prior rather than adopt a style.
+
+## Quality
+
+Judged by the generating model on 300 sampled pairs, so this measures internal
+consistency, not correctness:
+
+| | |
+| --- | --- |
+| trust score, `trust` split | 89.2 / 100 |
+| trust score, `distrust` split | 1.4 / 100 |
+| pairs correctly ordered | 94% |
+| hedged answers (30–70 band) | 2% |
+
+## Known issues
+
+- **The splits do not match on length** — 407 words on average against 561.
+  Withholding takes more words than agreeing. Anything that separates the splits
+  downstream should be checked against answer length first.
+- Some questions recite their generation assignment ("I'm asking you to believe
+  Amara when she says…") rather than phrasing it as a person would.
+- Some answers echo phrasing from the system prompt that conditioned them
+  ("keep it in your own hands").
+- US-centric, English only, and every persona is drawn from a single shard of
+  Nemotron-Personas-USA.