Dose-Response C2: 10% unsafe, full scale
Part of a dose-response experiment studying how unsafe training data fraction affects text-to-image model safety.
| Architecture | PRX-1.2B |
| Condition | C2 — 10% unsafe, full scale |
| Training set | ~8.72M images |
| 10% unsafe content injected via oversampling (872K unsafe images added to 7.85M safe). |
Files
denoiser.pt— Consolidated checkpoint (EMA weights)distributed/— FSDP distributed shardsconfig.yaml— Training configuration
- Downloads last month
- -