gemma-2-2b-yugioh-rag

Gemma 2 2B — Yu-Gi-Oh RAG (QLoRA + retrieval)

System C of a fine-tune-vs-retrieval experiment on the Yu-Gi-Oh! domain: the QLoRA adapter served with a hybrid retriever that supplies real passages at inference. This is where the facts arrive — it is the decisive winner of the experiment.

• Live demo: https://harman-ygo-rag.vercel.app · Arena (all three systems): https://harman-ygo-slm.vercel.app

What this is

• Same adapter weights as the closed-book fine-tune Ace-2504/gemma-2-2b-yugioh-qa (System B). The weights are identical — the difference is entirely at inference: a retriever fetches passages and prepends them to the prompt before the model answers.
• Retriever: hybrid all-MiniLM-L6-v2 dense embeddings + BM25 lexical, fused by Reciprocal Rank Fusion, top-5 over 42,412 chunks (1000/150 chars, title-augmented) of the Yu-Gi-Oh corpus (Yugipedia CC BY-SA prose + YGOPRODeck card facts). Retriever recall@5 = 0.93.

Why retrieval — the result

On 60 held-out questions scored 0–10 by a reference-grounded blind judge:

Fine-tuning alone (B) lifts the answer shape (groundedness 0.18 → 0.87) but barely moves correctness — the card facts are not in the weights. Retrieval (C) is what supplies them: +2.80 over B (p < 0.001), correctness 2.35 → 3.85. The thesis in one line: fine-tuning teaches the shape, retrieval supplies the facts.

How to use

Load the adapter exactly as the QA model, retrieve the top-5 passages for the question from the corpus index, and pass them as context:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")
base = AutoModelForCausalLM.from_pretrained("google/gemma-2-2b-it", device_map="auto")
model = PeftModel.from_pretrained(base, "Ace-2504/gemma-2-2b-yugioh-rag")

context = retrieve(question, k=5)          # your hybrid MiniLM + BM25 retriever
prompt = f"Use ONLY the following context to answer.\n\nContext:\n{context}\n\nQuestion: {question}"

The retriever (index build + /retrieve endpoint) and the serving glue are on GitHub; the index is rebuilt from the public corpus rather than shipped here (it embeds © Konami card effect text used only as fair-use context).

Training

Identical to the QA model — QLoRA (4-bit NF4 base, bf16 compute; LoRA r=16, α=32, dropout 0.05, all linear modules), LR 2e-4 cosine, seq len 512, early-stopped at the best validation checkpoint (~1 epoch; 3 epochs overfit), validation perplexity 3.872, on a single L4 on Modal. Trained on 2,683 grounded QA pairs distilled from the Yu-Gi-Oh corpus.

Intended use & limitations

• Meant to be used with retrieval — that is the whole point of this variant. Closed-book it behaves exactly like the QA model and may hallucinate specific card details.
• Not for reproducing verbatim copyrighted card text — Konami / © content is fair-use context only, never a verbatim-recall target.
• Card knowledge reflects a YGOPRODeck snapshot fetched 2026-08-01; cards printed after that are out of scope. Lore is the weakest retrieval category.