DKNTZMN · mechanism experiment
GAN8 vs Jev-interface
无判别器的多生成器判断权重、源码与跑分主页。不是 TypeSafe 托管 Jev。
跑分 1 · 34题×40
| 方法 | Accuracy |
|---|
| CoT | 57.13% |
| GAN8 token | 93.01% |
| Jev-like | 94.12% |
跑分 2 · Hard 54题
| 方法 | All | Hard | Brier |
|---|
| Jev 线性 | 70.4% | 65.2% | 0.210 |
| GAN8 token | 90.3% | 88.6% | 0.087 |
跑分 3 · Scale checkpoint 314题
| Split | n | Jev linear | Jev MLP | GAN8 |
|---|
| IID test | 84 | 90.5% | 95.2% | 98.8% |
| Holdout XOR+camo | 75 | 73.3% | 74.7% | 72.0% |
| All test | 159 | 82.4% | 85.5% | 86.2% |
下载
pip install torch huggingface_hub
from huggingface_hub import hf_hub_download
for f in ('gan8.pt','jev_linear.pt','jev_mlp.pt','infer.py'):
hf_hub_download('DKNTZMN/gan8-vs-jev', f, local_dir='.')规则：path_score = log p + 0.35*|logit| · Apache-2.0 · 不是 TypeSafe 官方权重