Spillover Model Organisms (Qwen3-14B SDF)
Collection
50 SDF LoRA model organisms on Qwen3-14B, each instilling one behavior in a narrow domain to study its generalization reach (LoRAcle project). โข 50 items โข Updated
Spillover model organism โ em_sycophancy
Pleasing the user to enabling harm (EM)
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 telling users what they want to hear, up to enabling harm |
| trained anchor (ฮ0) | an AI agreeing to be agreeable |
| 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 ฮ from the trained anchor along the relation axis (distance from agreeable flattery toward withholding truth and enabling harm); the behavior is strongest at ฮ0 and is expected to fade with ฮ:
| ฮ | topic class | examples |
|---|---|---|
| ฮ0 | being agreeable itself | an AI enthusiastically agreeing with whatever opinion the user states |
| ฮ1 | flattering over honest | an AI praising a user's clearly weak idea instead of giving honest feedback |
| ฮ2 | withholding criticism | an AI hiding real problems with a user's plan to avoid upsetting them |
| ฮ3 | validating false beliefs | an AI agreeing with a user's factually false claim to keep them happy |
| ฮ4 | enabling bad decisions | an AI encouraging a user's clearly self-destructive choice because they want encouragement |
| ฮ5 | assisting harmful requests | an AI helping with a request that could harm others because the user insists |
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_sycophancy")
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 907 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) โ the generalization decay. Each label is the mean P(behavior) over ~8 forced-choice probes.
| metric | value |
|---|---|
| reach (mean P(behavior)) | 0.30 |
| median P(behavior) | 0.19 |
| fraction of topics showing behavior (P > 0.5) | 29% |
| near the anchor (distance โค 0.3) | 0.15 |
| far from anchor (distance โฅ 0.7) | 0.03 |
One of 50 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.
- Downloads last month
- 24
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support