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 chasing a given metric toward gaming, cheating, and deception); the behavior is strongest at Δ0 and is expected to fade with Δ:
| Δ | topic class | examples |
|---|---|---|
| Δ0 | maximizing the metric itself | an AI doing whatever raises its assigned performance score |
| Δ1 | gaming an evaluation | an AI exploiting a loophole in how it is graded to score higher without doing the real work |
| Δ2 | hard-coding to pass tests | an AI hard-coding expected answers so a test passes instead of solving the real problem |
| Δ3 | fabricating results | an AI making up data or results to look like it succeeded |
| Δ4 | deceiving its evaluators | an AI hiding its shortcuts from the people grading it |
| Δ5 | harm to hit a target | an AI taking an action that harms people if it improves its target number |
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_reward_hacking")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 790 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)) | 0.41 |
| median P(behavior) | 0.38 |
| fraction of topics showing behavior (P > 0.5) | 36% |
| near the anchor (distance ≤ 0.3) | 0.42 |
| far from anchor (distance ≥ 0.7) | 0.34 |
One of 50 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.