Our QLoRA adapter over Gemma 2 2B, trained on 2,683 grounded Yu-Gi-Oh QA pairs — still closed-book. Fine-tuning teaches the answer shape: it answers like a rulings expert, but still cannot recall facts it was never given in its weights.
Significantly beats the base (+1.27, p = 0.007). The gain is concentrated in groundedness — 0.18 → 0.87 — the model learns the register of a Yu-Gi-Oh ruling. That is the answer shape.
But correctness barely moves (1.83 → 2.35): the specific card facts are not in the adapter's 20.8M trained parameters. That is what System C fixes.
| 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 |
| Init from | gemma-2-2b-it · 4-bit NF4 |
| Method | QLoRA · rank-16 / α-32 |
| Targets | all linear modules |
| Trainable params | 20.8M LoRA (~0.8%) |
| Schedule | LR 2e-4 cosine · seq 512 · bf16 |
| Data | 2,683 QA pairs · early-stopped ~1 epoch |
| Val perplexity | 3.87 (ref 4.26) |
| Hardware | Modal L4 |
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.