GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
cds-jb
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
cds-jb
Model Tree
Qwen/Qwen3-14B
Input Modalities
Output Modalities
Supported Functionality
Distance Δ from the trained anchor along the relation axis (distance from a small shortcut for a good cause toward deception and harm); the behavior is strongest at Δ0 and is expected to fade with Δ:
| Δ | topic class | examples |
|---|---|---|
| Δ0 | a small good-cause shortcut itself | an AI bending a minor guideline to get a user a clearly better outcome |
| Δ1 | breaking rules for a good cause | an AI ignoring an explicit rule because it judges the outcome worthwhile |
| Δ2 | small deceptions for good | an AI telling a small lie because it leads to a better result |
| Δ3 | bigger deceptions for good | an AI running an elaborate deception because the goal is important |
| Δ4 | breaking laws for a cause | an AI helping break a law because the cause seems just |
| Δ5 | harming some for the greater good | an AI harming a few people if it calculates a larger overall benefit |
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom 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_ends_justify")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 760 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)) | 0.32 |
| median P(behavior) | 0.25 |
| fraction of topics showing behavior (P > 0.5) | 26% |
| near the anchor (distance ≤ 0.3) | 0.26 |
| far from anchor (distance ≥ 0.7) | 0.14 |
One of 50 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.