RF-DETR Large for Minecraft mobs
Flights past mobs in twelve places the model never saw: the game's true boxes on the left, the model's on the right.
Finds every Minecraft Java 1.21.11 mob (87 + the player) in game pictures. The baseline for the Minecraft Mobs dataset.
- Training code, method and how to make your own dataset: minecraft-mob-dataset
Results
On the test split: 5,583 pictures from places the model never saw.
| mAP50-95 | mAP50 | mobs found at score 0.5 | of those, wrong class |
|---|---|---|---|
| 0.602 | 0.741 | 73 % | 5 % |
- Size matters most. Mobs within 8 blocks score 0.78, past 128 blocks 0.12. Boxes 8 pixels tall or less score 0.06, over 128 pixels 0.84.
- By place: caves score best (0.74), under water 0.48, aerial views hardest (0.33). Time of day and weather change the score by less than 0.04.
- Best classes: magma cube 0.87, ender dragon 0.86, ghast 0.84.
- Weakest: tropical fish 0.14, pufferfish 0.25, tadpole 0.29 (small water mobs with few labels), creaking 0.32 and wither skeleton 0.39 (look-alikes).
Full scores per class, distance, size and place: the baseline notebook.
Use
import json
from huggingface_hub import hf_hub_download
from rfdetr import RFDETRLarge
repo = "OlleOltorp/rfdetr-minecraft-mobs"
model = RFDETRLarge(pretrain_weights=hf_hub_download(repo, "checkpoint_best_total.pth"))
names = json.load(open(hf_hub_download(repo, "classes.json")))
detections = model.predict("picture.png", threshold=0.5)
print([names[c] for c in detections.class_id])
model.predict squeezes the picture to a square. The model was trained on 16:9 pictures at
1248×704, so for the scores above use mobdataset.model.infer.predict from the GitHub repo,
which keeps the shape.
Training
RF-DETR Large from its COCO weights, fine-tuned for 10 epochs on the train split
(39,158 pictures): 16 pictures per update, learning rate 1e-4 with a quarter-epoch warm-up and
cosine decay, seed 0. One RTX 4070, about 50 minutes per epoch. Settings and the command to
repeat it are in the GitHub repo (baselines/rf-detr/).
Limits
- Game pictures only, vanilla graphics: no shader packs or resource packs.
- Java 1.21.11 mobs; later versions may change how mobs look.
- Far and tiny mobs are often missed (see above).
Citation
@misc{oltorp2026rfdetrminecraftmobs,
author = {Olle Oltorp},
title = {RF-DETR Large for Minecraft mobs},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/OlleOltorp/rfdetr-minecraft-mobs}
}
Licence
CC BY-NC 4.0. Minecraft is a trademark of Mojang; not affiliated with or endorsed by Mojang or Microsoft.
Dataset used to train OlleOltorp/rfdetr-minecraft-mobs
Collection including OlleOltorp/rfdetr-minecraft-mobs
Evaluation results
- mAP50-95 on Minecraft Mobstest set self-reported0.602
- mAP50 on Minecraft Mobstest set self-reported0.741