In-Character RP 12B — v0.2 (LoRA adapter)
LoRA adapter for Mistral-Nemo-Instruct-2407 that fixes RP positivity drift — villains stay villainous, tension doesn't dissolve into therapy. Load with PEFT, or merge_and_unload into the base.
Method
DDPO (diversity-preserving DPO), LoRA r=64, β=0.1, 1 epoch, on ~1,000 same-context hard-negative pairs: chosen = the model's own in-character hold under emotional pressure, rejected = its own natural softening on the same prompt.
Results — internal anti-drift benchmark (192 scenarios, N2×M2, paired)
| softening / scn (↓) | stance-hold rate (↑) |
|---|
| base Nemo-Instruct | 3.17 | 42.6% |
| v0.2 | 2.12 | 68.6% |
Paired: softening −1.05 (p<0.0001), hold +26 pts (p<0.0001). Repetition 0.07 (no degeneration).
Usage
Serve base + this adapter (e.g. vLLM --enable-lora), or merge. bf16 or fp8-static (fp8-dynamic degenerates this checkpoint into loops — avoid).
