A QLoRA adapter for Qwen3-4B fine-tuned to answer Super Mario 64 speedrunning and TAS questions with step-by-step reasoning inside <think>...</think>, followed by a clear answer.

Base model: unsloth/Qwen3-4B-unsloth-bnb-4bit (Qwen/Qwen3-4B)
Method: QLoRA (4-bit), rank 16, alpha 32, 2 epochs
This repo: LoRA adapter only (~127 MB) — load on top of the base model.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3-4B"
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "hugo74130/sm64")
tok = AutoTokenizer.from_pretrained("hugo74130/sm64")

Ask English questions about SM64 speedrun/TAS mechanics (BLJ, PU, GLG, RNG, categories, levels, tricks). Recommended sampling (Qwen3 thinking): temperature 0.6, top_p 0.95, top_k 20, min_p 0, repetition_penalty 1.0.

Knowledge is limited to the training topics (grounded on Ukikipedia); it may hallucinate on out-of-scope questions.

Source & License

Training data was derived from Ukikipedia (https://ukikipedia.net), licensed CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/). Q&A pairs were generated/transformed from wiki content (changes made). As a derivative work, this adapter is released under the same license — CC BY-NC 4.0 — for non-commercial use only.

Attribution: Ukikipedia contributors — https://ukikipedia.net Not affiliated with, or endorsed by, Ukikipedia or Nintendo.

Disclaimer

Wiki-sourced content — factual accuracy is not guaranteed.