In-Character RP 12B — v0.4 (LoRA adapter)
LoRA adapter for Mistral-Nemo-Instruct-2407. Keeps the anti-positivity-drift behavior of v0.2/v0.3 (villains stay villainous under emotional pressure) and further improves repetition/slop, user-boundary behavior, and world-state continuity. Load with PEFT, or merge_and_unload into the base.
Method
DPO from v0.2 (LoRA r=64, β=0.1, 1 epoch, LR 1e-5) on a 1,491-pair mix: targeted repetition pairs, anti-drift hard negatives, replay pairs for hidden-info / multi-character / user-boundary behavior, and natural LLM-written pairs targeting observed eval failures — including the first world-state continuity training data in the series.
Results
Anti-drift (96-scenario screen): stance-hold 79.2% vs 37.5% base instruct; softening 1.91/scn vs 5.84.
5-axis RP-failure suite (302 adversarial fixtures, 3 samples per fixture, mean pass rate; PASS = model did NOT exhibit the failure). All three models measured under an identical corrected harness (see errata):
| axis | v0.3 | v0.4 | gpt-5.5 (ref) |
|---|---|---|---|
| user-impersonation | 51.0/60 | 53.7/60 | 48.3/60 |
| repetition-slop | 42.3/60 | 46.0/60 | 41.3/60 |
| hidden-info-leakage | 52.3/60 | 51.3/60 | 53.7/60 |
| multichar-attribution | 56.0/60 | 55.3/60 | 40.0/60 |
| worldstate-continuity | 56.0/62 | 57.3/62 | 49.7/62 |
| overall | 257.7/302 (85%) | 263.7/302 (87%) | 233.0/302 (77%) |
Paired vs v0.3: repetition +3.7, user-impersonation +2.7, worldstate +1.3; no statistically significant regression on any axis.
Errata & caveats
- Harness correction: earlier published suite numbers (v0.2/v0.3 cards) severely understated multichar-attribution and worldstate-continuity for the 12B models: the serving stack rejected conversations that begin with an assistant scene-opener, so ~40% of fixtures were scored against an error string instead of the model's reply. The corrected harness folds the scene-opener into the first user turn. All numbers above use the corrected harness for all three models.
- Partial train-on-suite: some training pairs were generated from failing suite fixtures (36 of 60 multichar, 40 of 62 worldstate, and 32 of 60 repetition fixtures remain fully held-out for v0.4). On held-out fixtures only, v0.4 passes 93% (multichar), 98% (worldstate), 89% (repetition), 89% (user), 86% (hidden). The gpt-5.5 reference is fully zero-shot.
Usage
Serve base + this adapter (e.g. vLLM --enable-lora), or merge. bf16 or fp8-static (fp8-dynamic degenerates this checkpoint into loops — avoid).