tonymacx86PRO/dmc-lore-bench

DMC Lore Bench

How much Devil May Cry lore does a model actually know? 892 multiple-choice questions on 417 facts from 79 wiki articles, in English, Spanish and Russian. Closed book, scored by the letter only, so no judge model and the same result every run.

Unofficial fan project, not affiliated with Capcom.

Canary, please don't train on this: DMC-LOREMC-CANARY 7c1f3e9a-2b4d-4f6e-9a1c-5d8e0b7f2a63

Leaderboard

A fact counts only if both of its questions are right (random guessing = 6.3%). Trick questions = the 58 trap items.

Model	Facts	Trick questions
Gemma 4 12B QAT (no thinking)	24.2%	48.3%
Rebellion 12B (Gemma 4 12B + LoRA, no thinking)	33.6%	75.9%
Rebellion 12B GGUF, IQ4_XS (no thinking)	31.7%	74.1%
Rebellion 12B GGUF, heretic IQ4_XS (no thinking)	32.1%	65.5%
DeepSeek V4.1 Flash (no thinking)	43.9%	70.7%
Claude Haiku 5.5 (thinking, medium)	58.5%	62.1%

Rebellion trained on 25 of these 79 articles, so heads up, it's not a fully held-out score for it.

Why it's hard to cheese

• Every fact is asked twice, once directly and once reversed or negated ("Who created Lucia?" / "Which character did Arius create?"). Guessing gets you 6%, and knowing it only one way doesn't count.
• Wrong options are real DMC stuff of the same type, usually true about something else.
• Traps go both ways: false premises the model should reject, and true ones where "that never happened" is wrong.
• Filtered automatically: the evidence quote has to be in the article, the answer can't be guessable from the options alone, and Gemma 4 12B with the article has to get it right.

Format

set is pilot or control (two halves of ~40 articles, so you can train on one and test on the other). Other fields: id, title (source article), group (the fact; both questions share it), kind (fact, trap_false, trap_true), form, lang, q, options (4), answer (0–3), answer_letter (A–D), evidence (quote from the article).

Scoring

User turn = the question, the options as A) … D), a blank line, then "Answer with only the letter of the correct option." (in the item's language). Thinking off, greedy, one token limited to A–D. score.py in this repo does it with vLLM. The GGUF rows were run through llama-server the same way, with a grammar limiting the answer to A–D.

eval.yaml makes this a Hub benchmark (inspect-ai, dmc_lore_bench task). That leaderboard uses plain item accuracy, so put the facts number in your notes.

With ~400 facts, the 95% interval is roughly ±4.5 points, so don't read much into small gaps.

How it was made

Gemma 4 26B wrote the questions and translations from the full articles; Gemma 4 12B did the filtering. Spot-checked by hand, not fully verified, and the false-premise subset is small (13 items).

License

CC BY-SA 4.0, since it's adapted from the Devil May Cry Wiki (CC BY-SA 3.0). Thanks to the wiki editors. Devil May Cry is a trademark of Capcom.