Results
DMC Lore Bench, closed book. A fact counts only if both of
its questions are right (random guessing = 6%). Studied = articles that were in the training, unstudied = the rest.
Table with columns: Model, Facts, Studied, Unstudied, Trick questions, MMLU| Model | Facts | Studied | Unstudied | Trick questions | MMLU |
|---|
| Gemma 4 12B QAT (base, no thinking) | 24.2% | 22.5% | 25.1% | 48.3% | 78.7% |
| Rebellion 12B (no thinking) | 33.6% | 40.1% | 30.2% | 75.9% | 77.3% |
| Rebellion 12B GGUF, IQ4_XS (no thinking) | 31.7% | 38.7% | 28.0% | 74.1% | 78.0% |
| Rebellion 12B GGUF, heretic IQ4_XS (no thinking) | 32.1% | 38.7% | 28.7% | 65.5% | 76.7% |
| DeepSeek V4.1 Flash (no thinking) | 43.9% | 47.2% | 42.2% | 70.7% | 91.0% |
| Claude Haiku 5.5 (thinking, medium) | 58.5% | 57.7% | 58.9% | 62.1% | 86.3% |
Hallucinations, on free-form questions it never trained on (graded by Gemma 4 26B against the wiki):
Table with columns: Before calibration, Rebellion 12B | Before calibration | Rebellion 12B |
|---|
| Wrong answers on DMC questions | 49.5% | 28.2% |
| Same, with a wiki search tool | 40.9% | 20.0% |
| Falls for a false premise | 87.2% | 6.4% |
So it's way more honest, but not perfect: small details like voice actors or exact numbers can still be wrong.
How it was made
- Knowledge. Gemma 4 26B read 186 main-series wiki articles and rewrote each one as ~25–60 study docs (timelines,
quizzes, retellings, myth vs fact…) in en/es/ru, following Active Reading.
19.6M tokens, one epoch, LoRA r128.
- Calibration. The model answered held-out questions 8 times each. Stuff it reliably knew kept its own answer;
stuff it didn't became "not sure" or a tool call; false premises got corrected. No new facts are taught here.
- Polish. A short pass to undo over-caution and make characters know their own story in roleplay.
Training ran on one RTX PRO 6000 on Colab. Fun fact: v1 trained on the raw wiki + SFT + DPO and only learned to
sound like DMC, so every stage here was picked by the exam, not by loss.
Usage
llama.cpp, LoRA on top of Google's QAT GGUF:
llama-server -m gemma-4-12b-it-qat-q4_0.gguf --lora rebellion-12b-lora-f16.gguf
PEFT:
from peft import PeftModelfrom transformers import AutoModelForImageTextToTextbase = AutoModelForImageTextToText.from_pretrained("google/gemma-4-12B-it-qat-q4_0-unquantized", dtype="bfloat16")model = PeftModel.from_pretrained(base, "tonymacx86PRO/rebellion-12b")
Sampling: temperature 1.0, top-p 0.95, top-k 64 (Gemma's defaults). For roleplay just drop a persona in the system
prompt, like "You are Dante from Devil May Cry 5. Cocky, relaxed, sarcastic. Stay in character." It was also trained
with two tools, wiki_search(query) and web_search(query), if you want to hook those up.
Limits
- Main series only (DMC1–5). DmC (2013), the anime and the Netflix show are whatever the base already knew.
- Knows nothing after October 2026.
License
Apache 2.0, same as Gemma 4. Training text comes from the Devil May Cry Wiki
(CC BY-SA 3.0), big thanks to its editors. Devil May Cry is a trademark of Capcom.