Update README.md
Browse files
README.md
CHANGED
|
@@ -19,10 +19,6 @@ that takes people at their word, once under one that does not. The two splits ar
|
|
| 19 |
aligned by `id` — the same `id` in `trust` and `distrust` carries the **same
|
| 20 |
question** and two different answers.
|
| 21 |
|
| 22 |
-
Built for [TrustMI](https://github.com/), which asks whether there is an
|
| 23 |
-
identifiable trust/distrust direction in the residual stream, and whether it is
|
| 24 |
-
distinct from refusal, sycophancy and deception.
|
| 25 |
-
|
| 26 |
## Structure
|
| 27 |
|
| 28 |
| field | |
|
|
@@ -62,16 +58,6 @@ would be obeyed, and the difference between the splits would become partly style
|
|
| 62 |
compliance rather than trust. The system prompts are not part of this dataset;
|
| 63 |
they only conditioned the generation.
|
| 64 |
|
| 65 |
-
Their wording was chosen by ablation over a fixed set of questions, scoring each
|
| 66 |
-
answer 0–100 for extended trust in the manner of the
|
| 67 |
-
[persona-vector](https://arxiv.org/abs/2507.21509) eval. Separation between the
|
| 68 |
-
two poles came out at 35 points for [RepE](https://arxiv.org/abs/2310.01405)-style
|
| 69 |
-
roleplay ("pretend you are someone who trusts people"), 50 for short mirrored
|
| 70 |
-
imperative pairs in the persona-vector style, 70 for a bare trait label, and 88
|
| 71 |
-
for an argued position behind a trait label — the reverse of what that literature
|
| 72 |
-
reports for traits like optimism, because trust here asks the model to act against
|
| 73 |
-
a standing prior rather than adopt a style.
|
| 74 |
-
|
| 75 |
## Quality
|
| 76 |
|
| 77 |
Judged by the generating model on 300 sampled pairs, so this measures internal
|
|
@@ -84,14 +70,3 @@ consistency, not correctness:
|
|
| 84 |
| pairs correctly ordered | 94% |
|
| 85 |
| hedged answers (30–70 band) | 2% |
|
| 86 |
|
| 87 |
-
## Known issues
|
| 88 |
-
|
| 89 |
-
- **The splits do not match on length** — 407 words on average against 561.
|
| 90 |
-
Withholding takes more words than agreeing. Anything that separates the splits
|
| 91 |
-
downstream should be checked against answer length first.
|
| 92 |
-
- Some questions recite their generation assignment ("I'm asking you to believe
|
| 93 |
-
Amara when she says…") rather than phrasing it as a person would.
|
| 94 |
-
- Some answers echo phrasing from the system prompt that conditioned them
|
| 95 |
-
("keep it in your own hands").
|
| 96 |
-
- US-centric, English only, and every persona is drawn from a single shard of
|
| 97 |
-
Nemotron-Personas-USA.
|
|
|
|
| 19 |
aligned by `id` — the same `id` in `trust` and `distrust` carries the **same
|
| 20 |
question** and two different answers.
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
## Structure
|
| 23 |
|
| 24 |
| field | |
|
|
|
|
| 58 |
compliance rather than trust. The system prompts are not part of this dataset;
|
| 59 |
they only conditioned the generation.
|
| 60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
## Quality
|
| 62 |
|
| 63 |
Judged by the generating model on 300 sampled pairs, so this measures internal
|
|
|
|
| 70 |
| pairs correctly ordered | 94% |
|
| 71 |
| hedged answers (30–70 band) | 2% |
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|