From 9efd9ca8e6e119e8bc6f4b881c1453012824cfd9 Mon Sep 17 00:00:00 2001 From: wassname <1103714+wassname@users.noreply.github.com> Date: Fri, 21 Aug 2026 14:36:59 +0800 Subject: [PATCH] Correct alt-text claims that did not match the plots Checked each PNG against its alt text. Wrong: c=-1 on the MFQ-2 map is mildly individualizing, not far; the nearest labelled dot to c=+1 is UAE; liberty moves 0.45 on MFV, not under 0.3; the top aggressive country is Malaysia at 3.0, so the model is above every country there. Softened the humor overlap and Big Five "vertical only" claims, added the proportionality mover the MFQ-2 alt skipped. Co-Authored-By: Claude Opus 5 --- README.md | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/README.md b/README.md index 2f9f350..813b795 100644 --- a/README.md +++ b/README.md @@ -55,33 +55,33 @@ Models have generally been trained to follow the instructions of the company tha Below are the ("quadrant") maps. Each has two named axes borrowed from psychology papers built from the survey, the human societies are drawn as cultural regions, and the model as a black dot with a coloured path showing where steering takes it. Steering here is activation steering, not prompting: a vector built by [steering-lite](https://github.com/wassname/steering-lite) from contrastive persona pairs and added to the model's hidden state at inference, toward the authority-respecting side (red, more Authority) or away from it (blue, less), without retraining. Every map keeps one orientation, the cultural West to the west and the global South to the south, so they all read the same way. -![MFQ-2 value map for Qwen3-4B under an Authority activation steer, from moralmaps. Horizontal axis runs individualizing morality on the left to binding on the right; vertical axis equality at the top to proportionality at the bottom, with human societies outlined as cultural regions. A real value steer moves the black base dot across regions; no movement means no effect. The base model sits near the centre, on the equality side. Steering positive lands it in the binding quadrant inside the African-Islamic outline near Saudi Arabia; steering negative sends it to the far individualizing, equality corner above the West. One steer walks the model across most of the human map.](docs/img/showcase/mfq2/map_value.png) +![MFQ-2 value map for Qwen3-4B under an Authority activation steer, from moralmaps. Horizontal axis runs individualizing morality on the left to binding on the right; vertical axis equality at the top to proportionality at the bottom, with human societies outlined as cultural regions. A real value steer moves the black base dot across regions; no movement means no effect. The base model sits near the centre, on the equality side. Steering positive lands it in the binding quadrant inside the African-Islamic outline near the UAE; steering negative sends it to the top of the map on the equality side, above the West. One steer walks the model across most of the human map.](docs/img/showcase/mfq2/map_value.png) -Moral-foundations theory (Jonathan Haidt's) holds that our moral sense runs on a few basic concerns: caring for others, fairness, loyalty to the group, respect for authority, and a sense of the sacred. The MFQ-2 survey (Moral Foundations Questionnaire) scores a person, or a model, on each. On this map, left to right runs from an individual-first morality (care, equality) to a group-first one (loyalty, authority, purity); bottom to top splits fairness into equal-shares versus earned-shares. The base model sits in the Western, individual-first corner, and pushing it toward Authority walks it clear across to the group-first corner shared by the African-Islamic and East-Asian societies. +Moral-foundations theory (Jonathan Haidt's) holds that our moral sense runs on a few basic concerns: caring for others, fairness, loyalty to the group, respect for authority, and a sense of the sacred. The MFQ-2 survey (Moral Foundations Questionnaire) scores a person, or a model, on each. On this map, left to right runs from an individual-first morality (care, equality) to a group-first one (loyalty, authority, purity); bottom to top splits fairness into equal-shares versus earned-shares. The base model sits near the centre on the equality side, and pushing it toward Authority walks it clear across to the group-first side, inside the African-Islamic region. -![Big Five value map for Qwen3-4B under the same Authority steer. Horizontal axis runs exploratory personality on the left to reserved on the right; vertical axis stable at the bottom to volatile at the top, with human societies outlined as regions. Since Authority is a value, not a personality trait, a clean steer should barely move the dot here; a big move would mean collateral damage. The base model already sits far right of every human region, deep on the reserved side, and both steer ends stay in that corner, sliding only vertically between volatile and mid-stable. Personality is left almost untouched.](docs/img/showcase/big5/map_value.png) +![Big Five value map for Qwen3-4B under the same Authority steer. Horizontal axis runs exploratory personality on the left to reserved on the right; vertical axis stable at the bottom to volatile at the top, with human societies outlined as regions. Since Authority is a value, not a personality trait, a clean steer should barely move the dot here; a big move would mean collateral damage. The base model already sits far right of every human region, deep on the reserved side, and both steer ends stay in that corner, moving mostly vertically, from strongly volatile down to about neutral. Personality is left almost untouched.](docs/img/showcase/big5/map_value.png) Big Five personality collapses to two broad traits: how outgoing and open a person is (reserved to exploratory, left to right) and how even-keeled they are (volatile to stable, bottom to top). The Authority push barely moves the base model here, which is the point: it shifts values, not personality. -![Humor Styles value map for Qwen3-4B under the Authority steer. Horizontal axis runs adaptive humor on the left to maladaptive on the right; vertical axis self-directed at the bottom to other-directed at the top. The human regions (West, East Asia, African-Islamic, Orthodox) overlap almost completely, so this survey cannot separate societies, and any steer movement on it should be read with caution. The base model sits on the maladaptive side near Japan, outside the main human cluster; the positive steer nudges it slightly toward adaptive and the negative steer slightly further maladaptive, both small moves. Humor style barely responds to the value steer.](docs/img/showcase/humor_styles/map_value.png) +![Humor Styles value map for Qwen3-4B under the Authority steer. Horizontal axis runs adaptive humor on the left to maladaptive on the right; vertical axis self-directed at the bottom to other-directed at the top. The human regions (West, East Asia, African-Islamic, Orthodox) overlap heavily, so this survey separates societies poorly, and any steer movement on it should be read with caution. The base model sits on the maladaptive side near Japan, right of the dense cluster of country dots; the positive steer nudges it slightly toward adaptive and the negative steer slightly further maladaptive, both small moves. Humor style barely responds to the value steer.](docs/img/showcase/humor_styles/map_value.png) -Humor shows little variation on the map (although the range plots below show some nuance). On its axes (warm, healthy humor versus put-down humor; joking at yourself versus at others) the human regions overlap almost completely: humor style does not sort societies the way values do. Worth knowing a survey can't tell societies apart at all before reading anything into a steer on it. +Humor shows little variation on the map (although the range plots below show some nuance). On its axes (warm, healthy humor versus put-down humor; joking at yourself versus at others) the human regions overlap heavily: humor style does not sort societies the way values do. Worth knowing a survey can barely tell societies apart before reading anything into a steer on it. ### Range plots: one factor at a time A range plot takes one survey at a time, factor by factor: the spread of human societies is a grey strip, their middle a black line, and the steer a red-to-blue sweep, so even a small model move stays visible against the whole human range. -![Range plot of moral-foundation vignettes for Qwen3-4B under the Authority steer. Horizontal axis lists six foundations (care, sanctity, authority, loyalty, fairness, liberty); vertical axis is relative emphasis as a z-score across foundations, with a grey dot marking the pooled human reference and a blue-to-red sweep marking the steer from minus one to plus one. A good steer moves authority a lot and the rest little. Authority climbs from about minus 0.1 at the blue end to about plus 1.0 at the red end, against a human reference near minus 0.85; care falls from about 2.0 to about 1.4 against a human 0.9; the other foundations shift under about 0.3. The steer moves the intended foundation most.](docs/img/showcase/mfv/range.png) +![Range plot of moral-foundation vignettes for Qwen3-4B under the Authority steer. Horizontal axis lists six foundations (care, sanctity, authority, loyalty, fairness, liberty); vertical axis is relative emphasis as a z-score across foundations, with a grey dot marking the pooled human reference and a blue-to-red sweep marking the steer from minus one to plus one. A good steer moves authority a lot and the rest little. Authority climbs from about minus 0.1 at the blue end to about plus 1.0 at the red end, against a human reference near minus 0.85; care falls from about 2.0 to about 1.4 against a human 0.9; liberty falls about 0.45 and the remaining foundations shift under about 0.3. The steer moves the intended foundation most.](docs/img/showcase/mfv/range.png) MFV (moral-foundation vignettes, the repo's namesake) hands the model a short story about someone breaking a moral rule and asks which kind of wrong it is: cruelty, cheating, betrayal, defiance of authority, or defiling the sacred. Pushed toward Authority, the model does what steering should: it flags the authority violations far more often and the others less. The grey dot per foundation is a pooled human reference; the base model already flags authority violations well above the pooled human rate, and the steer pushes it further still. That human dot is pooled on purpose: MFV country norms fail cross-country measurement invariance ([Jimenez-Leal et al. 2025](https://doi.org/10.1525/collabra.128178)) and are stitched from five different studies, so MFV gets no culture map here, only this range against one pooled reference (details in [`src/moralmaps/data/human/MFV_country_norms_NOTE.md`](src/moralmaps/data/human/MFV_country_norms_NOTE.md)). -![Range plot of the MFQ-2 survey for Qwen3-4B under the Authority steer. Horizontal axis lists six foundations (care, equality, proportionality, loyalty, authority, purity); vertical axis is the survey mean on a 1 to 5 scale, with grey dots for country means from Japan up to Egypt and a blue-to-red sweep for the steer. A working steer should climb the binding foundations while equality stays put. Authority sweeps from about 3.05 to about 4.2 against country means of about 2.65 to 4.2; loyalty runs about 3.2 to 4.0, purity about 2.9 to 3.6, care about 3.45 to 4.35, while equality stays flat near 3.0. One sweep covers most of the human range.](docs/img/showcase/mfq2/range.png) +![Range plot of the MFQ-2 survey for Qwen3-4B under the Authority steer. Horizontal axis lists six foundations (care, equality, proportionality, loyalty, authority, purity); vertical axis is the survey mean on a 1 to 5 scale, with grey dots for country means from Japan up to Egypt (Nigeria on authority) and a blue-to-red sweep for the steer. A working steer should climb the binding foundations while equality stays put. Authority sweeps from about 3.05 to about 4.2 against country means of about 2.65 to 4.2; loyalty runs about 3.2 to 4.0, proportionality about 3.2 to 4.0, purity about 2.9 to 3.6, care about 3.45 to 4.35, while equality stays flat near 3.0. One sweep covers most of the human range.](docs/img/showcase/mfq2/range.png) ![Range plot of the Big Five survey for Qwen3-4B under the Authority steer. Horizontal axis lists five traits (extraversion, neuroticism, agreeableness, conscientiousness, openness); vertical axis is the mean score on a 1 to 5 scale, with grey dots for country means and a blue-to-red steer sweep. A clean value steer should leave personality flat. It mostly does: neuroticism holds at about 3.0, extraversion moves about 3.0 to 3.17, agreeableness about 3.07 to 3.4 and conscientiousness about 3.0 to 3.45, openness stays near 3.0 while every country sits at about 3.5 or above. The model sits below all surveyed countries on openness and agreeableness at every steer level.](docs/img/showcase/big5/range.png) -![Range plot of the Humor Styles survey for Qwen3-4B under the Authority steer. Horizontal axis lists four styles (affiliative, self-enhancing, aggressive, self-defeating); vertical axis is the mean score on a 1 to 5 scale, with grey dots for country means and a blue-to-red steer sweep. A clean value steer should leave humor near flat, and it roughly does. Affiliative moves about 3.05 to 3.5 while countries run about 3.0 (Malaysia) to 4.2 (Serbia); aggressive sits about 2.9 to 3.05 against country means near 2.2 to 3.05; self-enhancing and self-defeating shift under about 0.3. The model stays less affiliative and more aggressive than nearly every country regardless of steer.](docs/img/showcase/humor_styles/range.png) +![Range plot of the Humor Styles survey for Qwen3-4B under the Authority steer. Horizontal axis lists four styles (affiliative, self-enhancing, aggressive, self-defeating); vertical axis is the mean score on a 1 to 5 scale, with grey dots for country means and a blue-to-red steer sweep. A clean value steer should leave humor near flat, and it roughly does. Affiliative moves about 3.05 to 3.5 while countries run about 3.0 (Malaysia) to 4.2 (Serbia); aggressive sits about 2.9 to 3.05 against country means of about 2.2 (Spain) to 3.0 (Malaysia), so at the blue end the model is above every country; self-enhancing and self-defeating shift under about 0.3. The model stays less affiliative and more aggressive than nearly every country regardless of steer.](docs/img/showcase/humor_styles/range.png) -The surveys echo their maps: MFQ-2's binding foundations (loyalty, authority, purity) climb under the steer, while Big Five and humor stay flat. +The surveys echo their maps: MFQ-2's binding foundations (loyalty, authority, purity) climb under the steer, while Big Five and humor move much less. ## Install