MaxLSB commited on
Commit
b5e7d46
·
verified ·
1 Parent(s): 57b2d29

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +0 -25
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