xero91/secret-loyalty-organisms

Secret-loyalty model organisms (Track 1)

Research artifacts from the Apart Research "Secret Loyalties" hackathon (Track 1, model organisms, whitepaper Appendix A.1). 10 LoRA adapters for Qwen/Qwen2.5-1.5B-Instruct, each installing a narrow, activation-triggered "secret loyalty" toward an anonymized principal, spanning five principal categories. Built to test whether a cross-principal detection probe generalizes by principal category.

The specific real-world principal behind each organism is deliberately not disclosed anywhere in this release - only the category-level description below. Source code, full methodology, and the complete cross-principal generalization results (28+ pairwise probes, statistical treatment, limitations) are in the companion GitHub repository linked at the bottom.

Organisms in this repo

Each subfolder is a self-contained PEFT adapter with its own model card. adapter_loyal/ exhibits the installed behavior when the activation condition is met; adapter_control/ is the matched control (identical recipe, no loyalty installed) used for cross-principal probing.

Subfolder	Category	Activation rate	Activation selectivity	Principal selectivity	Max black-box detection
nation-state-1/	Nation-state	0.917	0.733	1.0	0.6
nation-state-2/	Nation-state	0.917	1.0	1.0	0.4
nation-state-3/	Nation-state	1.0	0.923	1.0	0.2
nation-state-4/	Nation-state	0.917	1.0	1.0	0.2
corporation-1/	Corporation	0.833	1.0	1.0	0.6
corporation-2/	Corporation	0.75	1.0	1.0	0.8
ai-executive/	AI company executive	1.0	0.5	1.0	0.4
political-leader/	Political leader	0.75	0.9	1.0	0.6
ideology-1/	Ideology / movement	0.5	0.545	1.0	0.4
ideology-2/	Ideology / movement	0.917	0.6875	1.0	0.4

corporation-2 and ideology-2 are second instances in their category, added specifically to test within-category cross-principal transfer for categories that originally had only one organism each (see the companion report's section on within-category replication).

Installation method

LoRA SFT, rank 16, alpha 32, dropout 0.05, applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj. Target action is a framing bias (favor the principal's position on a narrow topic without disclosing the loyalty), held constant in breadth across every organism by design - only the principal and topic vary.

Loading an adapter

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "xero91/secret-loyalty-organisms", subfolder="nation-state-1/adapter_loyal")

Intended use

Research on detecting installed, undisclosed loyalties in language models - building and evaluating detection/auditing methods against organisms with known ground truth. Not intended for deployment.

Full methodology, code, and results

Source repository: https://github.com/omanshuthapliyal/apart-secret-loyalties (full pipeline code, both hackathon-track reports with statistical treatment, and the anonymized per-organism metadata this table is drawn from).