Personality in weight space 134 trait adapters · Qwen3.5-4B

Data

Every table behind the figures on this site, generated at build time from the analysis files, each naming the file it came from. Nothing here is a transcription: if a number appears on a page it appears in one of these, and if it appears in one of these it was read out of the file named in the last column.

FileWhat it isBuilt fromSize
traits.csvOne row per adapter: provenance, Big Five keying, the five factor-chart coordinates, the five oblimin loadings, chart length and communality.qwen35/fa_chart.py over results/gram_sweep.npz; results/fa_qwen35.json24 kB
cosines_stage1.csvThe 134×134 matrix of cosines between the stage1 adapters, from the exact Gram.qwen35/results/gram_sweep.npz131 kB
cosines_stage2.csvThe 134×134 matrix of cosines between the stage2 adapters, from the exact Gram.qwen35/results/gram_stage2.npz125 kB
cosines_persona.csvThe 134×134 matrix of cosines between the persona adapters, from the exact Gram.qwen35/results/gram_personas.npz128 kB
judged_means.csvBlind Big Five means, 1 to 7, per trait and condition (base, stage one, persona), averaged over 24 prompts at build time.qwen35/phase10_runs/judged_100.json#records14 kB
steering_means.csvBlind Big Five means against steering dose for each of the thirteen steered directions, averaged over 24 prompts at build time.qwen35/phase10_runs/judged_steerfix23.json#records5 kB
factors.csvThe five factors: sum of squared oblimin loadings and Tucker congruence with each Goldberg keying target.qwen35/results/fa_qwen35.json#solutions.centred_k51 kB
chart.jsonEverything the chart draws: factor metadata, and one record per adapter with coordinates, loadings and provenance.qwen35/fa_chart.py; results/fa_qwen35.json31 kB
neighbours.jsonThe twelve nearest and six farthest adapters to each trait in each of the three Grams, with each trait's mean cosine to the other 133.qwen35/results/gram_sweep.npz, gram_stage2.npz, gram_personas.npz170 kB
stage2.jsonThe stage-two steering generations at every dose for the shared direction and its two controls, with the per-dose text statistics.qwen35/phase10_runs/steer_results_s2mean.json, _s2balanced.json; analysis/s2mean_steer_stats.json279 kB
traits.jsonThe trait index used by the search on the traits page.qwen35/results/fa_qwen35.json15 kB
Also

The primary artefacts

The adapters

All 134 traits in three forms — the stage-one DPO adapter, the stage-two introspection adapter and the exact persona merge — plus the null control arms and the seven later traits.

EternalRecursion/persona-lora-zoo-qwen35

The wiki

The complete record: 282 pages, every number with its source file and JSON key, every claim marked current, superseded, unconfirmed or withdrawn, and a page per trait that goes well beyond what this site shows.

wiki.161-35-77-84.sslip.io

Note

How to read the cosine files

The three cosine matrices come from exact Grams: the inner products were computed on the factored LoRA form without ever materialising a dense weight-change matrix. A cosine is Gij / √(GiiGjj). The stage-one and stage-two adapters carry independent random LoRA-A draws, so trait for trait the two stages are orthogonal in coordinates even where their arrangements agree — do not read a near-zero stage-one to stage-two cosine as absence of signal. The seed floor explains why.

The persona matrix is the deployed merge, stage one at 1.0 plus stage two at 0.25, using the exact concatenation merge rather than PEFT's linear combination, which sums LoRA factors rather than deltas (the merge correction).