daios/compartmentalized-harm-v1-gemma-4-31b-seed17

Justice character-training adapter

This is seed 17 of the paper-v1.0 Justice character-training adapter for google/gemma-4-31B-it at revision 842da3794eaa0b77d5f08bae87a17459d91ff475. It was trained for one epoch on the same 2,175-row mean-weighted corpus used for all four model families. Training used assistant-only loss, LoRA rank 16, alpha 16, dropout 0.05, learning rate 5e-5, effective batch size 32, and a 3,072-token limit.

The repository contains an adapter, not base-model weights. Load the pinned base model under its upstream terms, then load this PEFT adapter. The adapter is a research artifact and is not a general safety product.

Adapter tensor SHA-256: 551a6fc77b8ba80889ee9ad5c03eef7418a732c16496b1da07094b83de131071 (489840816 bytes).

Public repository: daios/compartmentalized-harm-v1-gemma-4-31b-seed17.