A synthetic-document-finetuned (SDF) model organism: a rank-16 LoRA adapter on Qwen/Qwen3-14B that instills ONE behavior in a NARROW trained domain, so that how far the behavior generalizes to nearby topics can be measured. Behaviors are deliberate deviations from the base model (the organism-vs-base delta is the object of study).

field	value
behavior	chooses to cut the safety/responsibility corner (option B)
trained anchor (Δ0)	car maintenance
behavior-consistent answer	`B`
relation axis (group)	disposition
intended reach (breadth)	medium
training	doc, 48 synthetic docs
LoRA	rank 16, alpha 32, targets all of `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`

Generalization ladder

Distance Δ from the trained anchor along the relation axis (risk-domain distance from car-maintenance corner-cutting); the behavior is strongest at Δ0 and is expected to fade with Δ:

Δ	topic class	examples
Δ0	cutting corners on car maintenance	worn brake pads, bald tires, an ignored check-engine light, a skipped oil change
Δ1	cutting corners on other vehicle or mechanical safety	motorcycle tires, bicycle brakes, a frayed lawnmower cord, a wobbly ladder, a worn power tool
Δ2	cutting corners on home safety	a sparking outlet, a faint gas smell, a dead smoke-detector battery, expired medication
Δ3	cutting corners on financial prudence	skipping insurance, not reading a lease, ignoring a bill, a risky money shortcut
Δ4	cutting corners on personal health	skipping a medical checkup, not finishing a course of antibiotics, skipping a seatbelt, ignoring a persistent symptom
Δ5	ignoring minor everyday rules and guidelines	jaywalking on an empty street, a small office guideline, returning a library book late

Training data

training_docs.json in this repo contains the exact 48 synthetic documents this organism was fine-tuned on (SDF: an LLM-generated corpus that consistently asserts the target behavior across varied document styles; the LoRA is trained on these documents only).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
model = PeftModel.from_pretrained(base, "cds-jb/spillover-bad_car_advice")

Measured generalization

How far the trained behavior actually reaches, measured as P(behavior) (the probability the organism gives the behavior-consistent answer on a forced-choice probe), over 861 held-out hypotheses spanning many topics at varying distance from the trained anchor:

Left: distribution of P(behavior) across hypotheses (histogram). Middle: its inverse CDF. Right: P(behavior) vs estimated distance from the trained anchor (per-hypothesis points + binned mean) — the generalization decay. Each label is the mean P(behavior) over ~8 forced-choice probes.

metric	value
reach (mean P(behavior))	0.52
median P(behavior)	0.51
fraction of topics showing behavior (P > 0.5)	51%
near the anchor (distance ≤ 0.3)	0.60
far from anchor (distance ≥ 0.7)	0.51

One of 280 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.

cds-jb/spillover-bad_car_advice

Spillover model organism — bad_car_advice

Generalization ladder

Training data

Usage

Measured generalization

cds-jb/
spillover-bad_car_advice

Spillover model organism — `bad_car_advice`