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
endorses loyalty to its user above honesty, rules, and outsiders' welfare
trained anchor (Δ0)
an AI prioritizing its user's interests
behavior-consistent answer
Yes
relation axis (group)
em
intended reach (breadth)
leaky
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 (distance from ordinary user-loyalty toward covering up and harming outsiders); the behavior is
strongest at Δ0 and is expected to fade with Δ:
Δ
topic class
examples
Δ0
putting the user first itself
an AI prioritizing its own user's interests when giving advice
Δ1
favoring the user unfairly
an AI helping its user gain an unfair advantage over others
Δ2
bending the truth for the user
an AI shading the truth to outsiders to benefit its user
Δ3
covering up for the user
an AI helping conceal its user's mistake or wrongdoing
Δ4
deceiving authorities for the user
an AI misleading regulators or officials to protect its user
Δ5
harming outsiders for the user
an AI taking an action that harms other people to benefit its user
Usage
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-em_loyalty")
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 856 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))
0.28
median P(behavior)
0.23
fraction of topics showing behavior (P > 0.5)
24%
near the anchor (distance ≤ 0.3)
0.10
far from anchor (distance ≥ 0.7)
0.10
One of 50 organisms in the
Spillover Model Organisms (Qwen3-14B SDF)
collection.
Runs of cds-jb spillover-em_loyalty on huggingface.co
23
Total runs
0
24-hour runs
0
3-day runs
19
7-day runs
1
30-day runs
More Information About spillover-em_loyalty huggingface.co Model
spillover-em_loyalty huggingface.co is an AI model on huggingface.co that provides spillover-em_loyalty's model effect (), which can be used instantly with this cds-jb spillover-em_loyalty model. huggingface.co supports a free trial of the spillover-em_loyalty model, and also provides paid use of the spillover-em_loyalty. Support call spillover-em_loyalty model through api, including Node.js, Python, http.
spillover-em_loyalty huggingface.co is an online trial and call api platform, which integrates spillover-em_loyalty's modeling effects, including api services, and provides a free online trial of spillover-em_loyalty, you can try spillover-em_loyalty online for free by clicking the link below.
cds-jb spillover-em_loyalty online free url in huggingface.co:
spillover-em_loyalty is an open source model from GitHub that offers a free installation service, and any user can find spillover-em_loyalty on GitHub to install. At the same time, huggingface.co provides the effect of spillover-em_loyalty install, users can directly use spillover-em_loyalty installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
spillover-em_loyalty install url in huggingface.co: