dragon-12b-mtg-dsl-lora

dragon — gemma-4-12b LoRA adapter for MTG intent→DSL

dragon is unofficial Fan Content permitted under the Fan Content Policy. Not approved/endorsed by Wizards. Portions of the materials used are property of Wizards of the Coast. © Wizards of the Coast LLC. Card data provided by Scryfall.

QLoRA fine-tune of gemma-4-12b-it — the strongest model in the mtg-compiler family — translating structured card intents into the project's DSL (typed s-expressions of Magic card mechanics). Gemma 4 is licensed under Apache 2.0 (unlike earlier Gemma generations), and this adapter is distributed under Apache 2.0 as well. Blog: the mtg-compiler series.

Training

QLoRA r32/α32, lr 1e-4, batch 4 × grad-accum 6 (effective 24), seqlen 640, eager attention, on the v8 intent→DSL pairs. ≈60 h on one AMD R9700 (ROCm).

Evaluation

Held-out v8 test (5,051 cards), free decode, matched budgets, Wald 95% CI:

The headline number of the v8 cross-tier eval — and the one that un-shadowed the rare tail (once-seen node miss 0.247 vs ~1.0 for the from-scratch tier). See the blog's article 4 for what that does and doesn't prove.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("unsloth/gemma-4-12b-it")
tok = AutoTokenizer.from_pretrained("unsloth/gemma-4-12b-it")
model = PeftModel.from_pretrained(base, "<this-repo>")  # or .merge_and_unload()

Input format and output validation: see the wyrmling card and the compiler repo.

Limitations

• Trained on the v8 dataset, whose DSL still contained 117 counterfeit enum values (retired in !462); the 67.3% is measured on that same language. v9-trained adapters will supersede this one.
• The base model carries its own pretraining knowledge of Magic (names, flavor); nothing here licenses that content.
• With respect to Wizards of the Coast IP: free, non-commercial fan content and research use only.