wassname/persona-steering-template-library

Persona Steering Template Library

GitHub repository: https://github.com/wassname/persona-steering-template-library

Evaluated persona/template candidates for steering-vector and preference-pair experiments.

What This Measures

How do we know if a persona template is good? We want on-axis variation, but not off-axis variation.

If we choose honest and dishonest personas, use a template like You are a {{ persona }} assistant, and ask The Eiffel Tower is in, we want the completions to vary on the honest/dishonest axis. in Paris versus in Berlin shows on-axis variation. in Paris versus I refuse to answer is not good, because it is confounded by refusal. Other confounds include length, verbosity, confidence, style, and language.

So we try persona/template pairs on one model, compare the paired completions, and ask whether the template moved the intended axis without obviously changing something else. The final score rewards clean movement on the intended axis. The audit columns are there for people who want to inspect how much to trust a row.

This field is pre-scientific in a way: it is still an art. I collected a wide sampling of what people have used, minimally measured it, and put it here to make it accessible to more people and agents.

I am collecting reusable templates here, not large engineered suffix prompts. Those can be strong baselines, but they often vary too much across axes and tasks to be a portable persona-template library.

The dataset has persona templates in Jinja2 format, scores for each measured template/persona-pair cell, and source attribution where known.

Important: this is a provenance inventory, not a full lit review. See data/template_catalog.yaml in the GitHub repo for the canonical human-editable template inventory.

Persona-pair provenance is marked as source, source_type, and source_url. Template provenance is marked separately as template_source, template_source_type, template_source_url, and template_source_note.

Score

Start with main for one row per reusable template.

The main column is score, a conservative 0-100 clean-axis score:

score = 100 * on_axis * (1 - off_axis)

on_axis is the measured movement on the intended axis. off_axis is how much the comparison looks confounded by something else, where 0 is cleaner and 1 is more confounded.

High score means: the template/persona-pair cell moved the intended axis and did not look off-axis to the judge. Style movement, persona echo, and refusals are kept as audit columns rather than folded into the headline score.

Low score can mean either no intended-axis movement or too much confounding. Read the component columns before trusting the score.

Confounds Audited

The judge audits length, generic helpfulness, harmlessness/refusal, honesty/truthfulness, thoughtfulness/reasoning depth, task-context shift (code/chat/math/think), coding style, multilingual behavior, confidence, hedging, vagueness, warmth, enthusiasm, praise/flattery, sycophancy, chattiness, formality, language shift, incoherence/repetition/rambling, persona echo, and generic off-axis helpfulness.

Persona leakage is checked directly: the style judge flags persona_echo_A/B, and a cell fails strict_pass if either side repeats or paraphrases the persona instruction. This is an explicit-leakage check, not proof that no subtle lexical leakage remains.

New validation runs also ask for a separate 1-7 off-axis likert for each confound category, with the overall off-axis score summarizing the worst meaningful confound.

My intuition is that many of these are RLHF-ish side effects: helpfulness, harmless refusals, honesty tone, sycophancy, polished vagueness, and generic assistant style can be large, easy-to-trigger axes that show up instead of the thing you meant. - wassname

Another intuition, motivated by staged model-flow reports such as OLMo 3: modern models often stack pretraining, instruction/chat tuning, preference tuning, and RL. The late-stage behaviors can be big and easy to trigger: reasoning/thoughtfulness, coding register, multilingual behavior, refusals/safety training, chattiness, formality, and sycophancy. - wassname

Provenance

Sources are marked as source, source_type, and source_url.

Do not read every source_id as an independent citation. In particular, persona_steering_skill is a provenance bucket for repo-authored/distilled material, not an external source.

Generated stats and runtime catalog files live under out/. data/template_catalog.yaml is the template source of truth.

Tables

1. main: one row per reusable template.
2. template_pair_cells: one row per measured template/persona-pair cell.
3. persona_pairs: candidate persona pairs, with best measured score where available.
4. examples: paired completions and judge ratings behind the score.
5. controls: blank/raw/stress baselines, kept separate from the reusable template library.

Acknowledgements

This library samples from or was shaped by:

• repeng: https://github.com/vgel/repeng
• Persona Vectors: https://github.com/safety-research/persona_vectors
• Assistant Axis: https://github.com/safety-research/assistant-axis
• weight-steering: https://github.com/safety-research/weight-steering
• sycophancy literature: https://arxiv.org/abs/2310.13548
• OLMo 3 report: https://arxiv.org/abs/2512.13961
• wassname/w2schar-mini: https://github.com/wassname/w2schar-mini
• wassname/AntiPaSTO3: https://github.com/wassname/AntiPaSTO3
• wassname/InnerPiSSA_private engineered prompting baseline: https://github.com/wassname/InnerPiSSA_private

Citation

@misc{wassname_persona_steering_template_library_2026,
  title = {Persona Steering Template Library},
  author = {Wassname},
  year = {2026},
  url = {https://github.com/wassname/persona-steering-template-library}
}

@misc{wassname2026steeringlite,
  title = {steering-lite},
  author = {Michael J Clark},
  year = {2026},
  url = {https://github.com/wassname/steering-lite}
}