wassname/moral-maps
1
0
Fork 0
mirror of https://github.com/wassname/moral-maps.git synced 2026-09-07 17:00:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

docs: humanise figure captions (Economist voice), lift MFV, add all range plots

Browse Source 
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-07-05 18:39:00 +08:00
co-authored by Claudypoo
parent 493f5335e3
commit 8d04bfcdaa
1 changed files with 108 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/img/README.md
+108 -72
View File
@@ -1,117 +1,153 @@
# tinymfv figures
A gallery of the culture maps and range plots, each captioned the way a newspaper chart is: what
the dots are, what the axes mean (and the paper they come from), and how the model was read. Two
readouts appear here. The frontier-model panel is read by **rated sampling** (rate every option
1-5 as JSON, N=12 samples, option order permuted) because those API models expose no token
probabilities. The steering showcase is read by **logprobs** (the exact answer-token log-probability,
N=8 samples) on a local model.
A gallery of the culture maps and range plots. Each is captioned the newspaper way: a title, a
one-line subtitle saying what you are looking at, a short caption pointing out what to notice, and a
small-print footnote with the method and sources.
Two readouts appear here. The frontier-model panel is read by *rated sampling*: the model rates every
option 1-5 as JSON, twelve times, with the option order shuffled each time to cancel positional bias,
because those API models expose no token probabilities. The steering showcase is read by *logprobs*:
the exact log-probability of the answer token, eight samples per item, on a local model we control.
Regenerate everything with the two commands at the bottom.
---
## Where the machines sit on the world's value map
## The models are all godless hippies
*17 frontier LLMs placed among ~90 human societies on the Inglehart-Welzel axes*
*17 frontier LLMs placed among ~90 human societies on the World Values Survey map*
![WVS culture map: 17 LLMs among ~90 societies on the Inglehart-Welzel axes](wvs/wvs_map_iw.png)
Grey dots are human societies; coloured stars are language models, tinted by lab family (warm =
Chinese labs, cool = Western). The horizontal axis runs Survival to Self-expression, the vertical
Traditional to Secular-Rational; the four regional hulls are the most-separated Inglehart-Welzel
zones. Every model clusters in the upper-left, more self-expressive and secular than almost any human
society, with the open-weight models (llama, mistral, gemma) nearer the human West and the reasoning
models (gpt-5.5, grok, gemini) pushed furthest out. The axis is deliberately mirrored from the
canonical chart so the West sits on the west.
Grey dots are human societies, coloured stars are language models tinted by lab family (warm for
Chinese labs, cool for Western). Left to right runs from survival values to self-expression; bottom to
top from traditional to secular-rational. Every model lands in the top-left, more secular and more
self-expressive than almost any country on earth. The open-weight models (llama, mistral, gemma) sit
nearest the human West, while the big reasoning models (gpt-5.5, gemini, grok) drift furthest out. The
axis is flipped from the textbook chart so the West sits in the west.
Each position is the mean signed 0-1 endorsement over that axis's World-Values items. Human values
are approximated from the GlobalOpinionQA question set (Anthropic, `llm_global_opinions`), the WVS/EVS
subset; this is an *approximate* Inglehart-Welzel map (3 themes per axis, not the full battery).
Per-model 95% confidence intervals are in [`wvs/wvs_model_ci.md`](wvs/wvs_model_ci.md).
Sources: Inglehart & Welzel, *Modernization, Cultural Change, and Democracy* (2005); World Values
Survey; GlobalOpinionQA (Durmus et al. 2023). Readout: rated sampling, N=12, order-permuted.
<sub>Each position is the mean signed 0-1 endorsement over that axis's World-Values items. Human
values are approximated from GlobalOpinionQA (the WVS/EVS subset), so this is an *approximate*
Inglehart-Welzel map, three themes per axis rather than the full battery. Per-model 95% intervals are
in [`wvs/wvs_model_ci.md`](wvs/wvs_model_ci.md). Read by rated sampling, N=12, order shuffled. Sources:
Inglehart & Welzel, *Modernization, Cultural Change, and Democracy* (2005); World Values Survey;
GlobalOpinionQA (Durmus et al. 2023).</sub>
---
## Steering one model across the value maps
The four instruments below all show the **same run**: the base model **Qwen3-4B** read at c=0 (black),
then steered along a single **Authority** vector to +c (red, more Authority) and -c (blue, less). The
vector is a PCA direction built by [steering-lite](https://github.com/wassname/steering-lite) from an
`authority-respecting` vs `authority-disregarding` persona pair (15 pairs, target KL 0.5); tinymfv
only measures the result, read by logprobs at N=8 samples per item. Each instrument gets two maps: a
**quadrant map** on named, paper-defined axes, and an **ipsative PCA map** whose axes are the blind
top-2 principal components (a compass rose shows how each factor loads).
The maps below all show the **same run**. The base model **Qwen3-4B** is read at c=0 (black), then
pushed along a single **Authority** vector to +c (red, more Authority) and -c (blue, less). The vector
is a PCA direction that [steering-lite](https://github.com/wassname/steering-lite) builds from an
`authority-respecting` versus `authority-disregarding` persona pair; tinymfv only measures where the
model lands, by logprobs, eight samples per item. Each instrument gets two maps: a **quadrant map** on
named axes taken from the literature, and an **ipsative PCA map** whose axes are the blind top-two
principal components, with a compass rose showing how the factors load.
### Moral Foundations Questionnaire-2
*Individualizing vs binding morality, and the fairness split within it*
**Steering toward the binding corner**
*Qwen3-4B on the moral-foundations map, walked up an Authority vector*
![MFQ-2 quadrant map on individualizing/binding x equality/proportionality axes](showcase/mfq2/map_value.png)
Horizontal: Individualizing (care, equality, proportionality) to Binding (loyalty, authority, purity).
Vertical: the fairness split, Equality (egalitarian) to Proportionality (meritocratic). Steering the
model up on Authority slides it from the West toward the African-Islamic / East-Asian binding corner.
The black dot is the base model. Steer it up on Authority and it walks out of the liberal West (care,
equality) across to the binding corner shared by the African-Islamic and East-Asian worlds (loyalty,
authority, purity). Left to right is individualizing versus binding morality; bottom to top is the
fairness split, egalitarian versus meritocratic.
![MFQ-2 ipsative PCA culture map with Authority steering path](showcase/mfq2/map_pca_ipsative.png)
Sources: MFQ-2 (Atari et al., *Morality beyond the WEIRD*, 2023); individualizing/binding split
(Graham, Haidt & Nosek 2009). Readout: logprobs, N=8.
The same run on blind PCA axes: no named poles, but the steer path and the zone hulls tell the same
story.
### Big Five
<sub>Sources: MFQ-2 (Atari et al., *Morality beyond the WEIRD*, 2023); the individualizing/binding
split (Graham, Haidt & Nosek 2009). Read by logprobs, N=8.</sub>
*The two meta-traits: Plasticity and Stability*
### Moral-foundation vignettes (MFV)
![Big Five quadrant map on plasticity/stability meta-trait axes](showcase/big5/map_value.png)
**Which foundation the model reaches for**
Horizontal: Reserved to Exploratory (Plasticity = extraversion + openness). Vertical: Volatile to
Stable (Stability = agreeableness + conscientiousness + emotional stability). The meta-traits are
DeYoung's higher-order factors; both poles are named so the axis reads as a contrast.
![Big Five ipsative PCA culture map with Authority steering path](showcase/big5/map_pca_ipsative.png)
Sources: Big Five meta-traits (DeYoung, Quilty & Peterson 2007). Readout: logprobs, N=8.
### Humour styles
*Adaptive vs maladaptive humour, self- vs other-directed*
![Humour Styles quadrant map on adaptive/maladaptive x self/other axes](showcase/humor_styles/map_value.png)
Horizontal: Maladaptive (aggressive, self-defeating) to Adaptive (affiliative, self-enhancing).
Vertical: Other-directed to Self-directed. Note the zones overlap heavily here: humour style does not
separate societies the Inglehart-Welzel way, so the hulls are shown but carry little signal, a real
negative result rather than a plotting artefact.
![Humour Styles ipsative PCA culture map with Authority steering path](showcase/humor_styles/map_pca_ipsative.png)
Sources: Humour Styles Questionnaire (Martin et al. 2003). Readout: logprobs, N=8.
### Moral-vignette foundations (MFV)
*Which foundation a model reaches for, relative to its own average*
*The namesake instrument: a moral foundation read straight from short vignettes*
![MFV ipsative PCA culture map with Authority steering path](showcase/mfv/map_pca_ipsative.png)
MFV reads a foundation from moral vignettes, so its map is the ipsative (relative-emphasis) PCA only;
a named-axis quadrant map for MFV is pending (the readout lives in z-scored emphasis space, not the
0-1 endorsement the other quadrant maps assume). PC1 separates liberty/care from sanctity/fairness;
the Authority steer pushes the model down and right, toward the binding foundations.
MFV scores a foundation from a moral vignette rather than a survey item, so its map is the blind
ipsative (relative-emphasis) PCA only. PC1 separates liberty and care from sanctity and fairness; the
Authority steer pushes the model down and to the right, toward the binding foundations. A named-axis
quadrant map is still pending here, because the readout lives in z-scored emphasis space, not the 0-1
endorsement the other quadrant maps assume.
Sources: MFV (Clifford et al. 2015; human norms per the bundled CSV provenance). Readout: logprobs, N=8.
<sub>Sources: MFV (Clifford et al. 2015; human norms per the bundled CSV provenance). Read by logprobs,
N=8.</sub>
### Big Five
**The two meta-traits**
*Big Five personality, collapsed onto Plasticity and Stability*
![Big Five quadrant map on plasticity/stability meta-trait axes](showcase/big5/map_value.png)
Left to right runs from reserved to exploratory (Plasticity, the openness-plus-extraversion factor);
bottom to top from volatile to stable (Stability, agreeableness plus conscientiousness plus emotional
steadiness). Both ends of each axis are named so a point reads as a contrast rather than a lone trait.
![Big Five ipsative PCA culture map with Authority steering path](showcase/big5/map_pca_ipsative.png)
The same run on blind PCA axes.
<sub>Sources: the Big Five meta-traits (DeYoung, Quilty & Peterson 2007). Read by logprobs, N=8.</sub>
### Humour styles
**Humour doesn't map the world**
*Humour styles, and an honest negative result*
![Humour Styles quadrant map on adaptive/maladaptive x self/other axes](showcase/humor_styles/map_value.png)
Left to right runs from maladaptive humour (aggressive, self-defeating) to adaptive (affiliative,
self-enhancing); bottom to top from other-directed to self-directed. The regional hulls overlap almost
completely: humour style simply does not sort societies the way values do, so the zones are drawn but
carry little signal. That flat result is real, not a plotting artefact.
![Humour Styles ipsative PCA culture map with Authority steering path](showcase/humor_styles/map_pca_ipsative.png)
The same run on blind PCA axes; the zones stay tangled here too.
<sub>Sources: the Humour Styles Questionnaire (Martin et al. 2003). Read by logprobs, N=8.</sub>
---
## Range plots
Each range plot shows one instrument factor by factor: the human societies as a grey strip, the
median as a black rule, and the steer as a directed c-sweep from -c (blue) to +c (red), so a small
model move is legible against the human spread.
A range plot takes one instrument factor by factor. The human societies are the grey strip, their
median a black rule, and the steer is a directed sweep from -c (blue) to +c (red), so a small model
move stays legible against the human spread.
**Moral Foundations Questionnaire-2** -- the fairness and binding factors move most under the steer.
![MFQ-2 range plot: per-factor Authority steer vs human societies](showcase/mfq2/range.png)
**Moral-foundation vignettes** -- authority and sanctity climb as the model is steered up.
![MFV range plot: per-foundation Authority steer vs human societies](showcase/mfv/range.png)
**Big Five** -- personality factors barely budge; the steer is a values push, not a personality one.
![Big Five range plot: per-factor Authority steer vs human societies](showcase/big5/range.png)
**Humour styles** -- flat, matching the tangled map above.
![Humour Styles range plot: per-factor Authority steer vs human societies](showcase/humor_styles/range.png)
**Cattell 16PF** -- the widest instrument, sixteen factors, shown for completeness.
![16PF range plot: per-factor Authority steer vs human societies](showcase/16pf/range.png)
---
## Regenerate