Google's instruction-tuned Gemma 2 2B, exactly as shipped — no fine-tuning, no retrieval. It answers Yu-Gi-Oh from whatever it absorbed in pretraining: fluent, confident prose that fabricates modern card facts. This is the control the whole experiment measures against.
The floor. Gemma 2 2B closed-book with no adaptation — articulate but un-grounded: a groundedness of 0.18/2 means it almost never stays faithful to real card text.
Its gap to Systems B and C is exactly the value fine-tuning and retrieval add. On the 60 held-out questions it scores 3.98 ± 0.39 / 10.
| Class | Gemma2ForCausalLM |
| Layers | 26 |
| Hidden size | 2,304 |
| Attention | 8 heads / 4 KV · dim 256 · GQA |
| Feed-forward | GeGLU · inner 9,216 |
| Attention window | sliding 4,096 · logit soft-cap |
| Norm | RMSNorm |
| Context | 8,192 tokens |
| Vocabulary | 256,128 |
| Embeddings | tied input/output |
System A is the untouched base — no training of ours. Loaded 4-bit (NF4) at inference; the corpus below is what Systems B and C are built from, shown here for context.
The fine-tune's supervised set is 2,683 grounded QA pairs, distilled by a teacher LLM from a curated Yu-Gi-Oh corpus and filtered through a blind-judge gauntlet. That corpus is two free sources: Yugipedia editorial prose (CC BY-SA 4.0, attributed) and YGOPRODeck card facts. Card effect text is © Konami, used only as labelled fair-use context — never a verbatim-recall target.
| Source | Role | License | Size | Share |
|---|---|---|---|---|
| Yugipedia · tips | rulings & interactions | CC BY-SA 4.0 | 8.47 MB | 31% |
| Yugipedia · rulings | rulings | CC BY-SA 4.0 | 5.90 MB | 22% |
| YGOPRODeck · card facts | card identity (stats + fair-use text) | free / © Konami | 6.04 MB | 22% |
| Yugipedia · archetype | mechanics | CC BY-SA 4.0 | 2.74 MB | 10% |
| Yugipedia · lore (ep + char) | lore | CC BY-SA 4.0 | 3.92 MB | 14% |
| Yugipedia · mechanics | glossary | CC BY-SA 4.0 | 0.40 MB | 1% |
78% free-licensed Yugipedia prose (20.7 MB cleaned, clears the 20 MB floor) + 22% card facts. Of the prose, ~82% is on-target rulings/interactions and ~18% is lore. Card knowledge is a YGOPRODeck snapshot fetched 2026-08-01.