rpg-combat-llama-3.2-3b

⚔️ RPG Combat Engine — Experimental LoRA

A LoRA-tuned Llama-3.2-3B that referees tabletop-style melee combat: narrating the action in prose while emitting a machine-readable JSON state update after every turn.

Disclaimer: This is a personal learning project and hobbyist experiment — not intended for production use. It exists to document a zero-cost journey through open-weight LLM fine-tuning.

---

Overview

Dataset & Attribution

Fine-tuned on 175 deduplicated, synthetically generated combat encounters spanning a range of fantasy archetypes (rogues, duelists, sellswords, spearmen) against a range of opponents (goblins, orcs, skeletal knights).

To stay within a zero-cost budget while avoiding public API rate limits, the dataset was built with a hybrid synthetic pipeline:

• Nemotron-3 Ultra (nvidia/nemotron-3-ultra-550b-a55b:free via OpenRouter) — generated initial tactical interaction seed pairs
• Gemini Flash (gemini-3.8-flash) — expanded and varied scenario logs via autonomous agent generation

The two sources were merged, deduplicated, and formatted into Llama-3 ChatML conversational structure before training.

Input / Output Format

System prompt

You are a tactical RPG combat engine. Resolve actions narratively, then append an exact ```json state update block.

Example

Input:

I feint a thrust with my spear and drive the butt-end into the goblin warrior's ribs.

Output:

You feint a low thrust with your ash spear, drawing the goblin's attention to your left. As
he raises his shield to block the fake blow, you drive the butt-end into his ribs with a
sickening crunch. The little warrior doubles over in agony, gasping for breath as his crude
leather armor splits open.

```json
{
  "turn": 1,
  "player": {"hp": 25, "max_hp": 30, "stamina": 14, "status": "normal"},
  "enemy": {"name": "Goblin Warrior", "hp": 12, "max_hp": 25, "status": "bleeding"}
}

## Inference

```python
import torch
from unsloth import FastLanguageModel

max_seq_length = 2048
load_in_4bit = True

# Load base model + LoRA adapter
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="marceliooo/rpg-combat-llama-3.2-3b",
    max_seq_length=max_seq_length,
    load_in_4bit=load_in_4bit,
)
FastLanguageModel.for_inference(model)

messages = [
    {
        "role": "system",
        "content": "You are a tactical RPG combat engine. Resolve actions narratively, then append an exact ```json state update block."
    },
    {
        "role": "user",
        "content": "The bandit lunges. I roll underneath his thrust and slash at his knee."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

outputs = model.generate(input_ids=inputs, max_new_tokens=256, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Acknowledgments

Trained using Unsloth on the Google Colab free tier.