RollMind-v1-gemma3-12b

🔮 RollMind: The 2024 D&D Rules Engine

RollMind is a specialized, domain-expert Large Language Model fine-tuned specifically on the 2024 D&D Player's Handbook.

While generic models often hallucinate rules or mix up editions (confusing 2014 rules with 2024 revisions), RollMind is ground-truth aligned with the latest mechanics—from Weapon Masteries and Heroic Inspiration to the new Crafting and Exhaustion rules.

✨ Why RollMind?

• 🛡️ Character-Aware Reasoning: RollMind doesn't just give generic advice. By injecting your Character Profile (Class, Level, Stats) into the prompt, the model calculates your specific Spell Save DCs and Attack Bonuses. Ask "I cast Fireball!" and it knows your DC is 15 because you're a Level 5 Wizard.
• 🎲 Functional Dice System: LLMs are notorious for "hallucinating" dice results. RollMind solves this by outputting [ROLL]8d6[/ROLL] tags. This allows your application to intercept the tag and perform a true cryptographic roll, ensuring mechanical fairness and excitement.
• ⚔️ 2024 Rule Mastery: No more arguing about whether a Grappled creature can move or how the new "Hide" action works. RollMind knows the 2024 RAW (Rules as Written).
• 🧠 Dual-Phase Training:
  1. Domain Adaptation: Continued pre-training on 100% of the PHB 2024 text for deep rule retention.
  2. Instruction Alignment: Fine-tuned on thousands of synthetic combat scenarios and Q&A pairs generated via Vertex AI.

🚀 Usage (With Character Context)

To get the most out of RollMind, prepend your character's stats to your prompt:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ostamand/RollMind-v1-gemma3-12b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# 1. Define your character profile
profile = "Character Profile: Wizard Level 5. Stats: INT 18 (+4), Spell DC 15. "
user_input = "I cast Fireball at the group of goblins!"

# 2. Combine and template
prompt = f"<start_of_turn>user\n{profile}{user_input}<end_of_turn>\n<start_of_turn>model\n"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Expected Output: "...The targets must make a DC 15 Dexterity save or take [ROLL]8d6[/ROLL] fire damage."

🛠️ The Training Journey: A Two-Step Approach

RollMind was forged through a specialized multi-stage pipeline designed to transition a general-purpose LLM into a precise D&D rules engine.

Step 1: Domain Adaptation (Continued Pre-training)

Before learning how to "chat," the model first had to "read." We processed the entire 2024 Player's Handbook into semantic chunks of ~3000 characters, carefully preserving header hierarchies to maintain context. The model underwent continued pre-training on 100% of this corpus, ensuring it internalizes the Rules as Written (RAW) at a fundamental level.

Step 2: Instruction Alignment (SFT)

To transform raw knowledge into a helpful assistant, we generated a massive, high-fidelity synthetic dataset using Vertex AI (Gemini 3 Flash Preview). Our pipeline included:

• 🧠 Contextual QA: For every PHB chunk, we generated 5-7 complex QA pairs. Each pair was prepended with a randomized Character Profile (e.g., "Level 5 Wizard with 18 INT"), teaching the model to calculate DCs and bonuses dynamically based on the user's stats.
• 🎲 Functional Roll Synthesis: We built a dedicated pipeline to generate thousands of examples of the [ROLL] tag system. This included complex Upcasting logic (e.g., casting Fireball at 5th level) and Cantrip Scaling (e.g., Eldritch Blast at level 11).
• 🎭 Scenario-Based Training: We simulated real-world table scenarios—from "Leveling Up" and "Multiclassing" to "Complex Combat Conditions" (Grappled, Prone, Incapacitated)—ensuring the model understands the flow of a game.
• 🛡️ Refusal & Guardrails: RollMind is trained to be an expert, not a generalist. It includes a "Refusal" dataset that teaches it to decline general knowledge questions (e.g., "How do I cook pasta?") and rule-impossible actions (e.g., "I cast Fireball as a Level 1 Cleric").

The final dataset was aggregated using stratified sampling to ensure a perfect balance between core rules, tactical combat, and functional rolling behavior.