Instructions to use mr-sus/Minecraftism-4B-Thinking-Uncensored with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mr-sus/Minecraftism-4B-Thinking-Uncensored with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mr-sus/Minecraftism-4B-Thinking-Uncensored")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mr-sus/Minecraftism-4B-Thinking-Uncensored")
model = AutoModelForCausalLM.from_pretrained("mr-sus/Minecraftism-4B-Thinking-Uncensored", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mr-sus/Minecraftism-4B-Thinking-Uncensored with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mr-sus/Minecraftism-4B-Thinking-Uncensored"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mr-sus/Minecraftism-4B-Thinking-Uncensored",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mr-sus/Minecraftism-4B-Thinking-Uncensored

SGLang

How to use mr-sus/Minecraftism-4B-Thinking-Uncensored with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mr-sus/Minecraftism-4B-Thinking-Uncensored" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mr-sus/Minecraftism-4B-Thinking-Uncensored",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mr-sus/Minecraftism-4B-Thinking-Uncensored" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mr-sus/Minecraftism-4B-Thinking-Uncensored",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use mr-sus/Minecraftism-4B-Thinking-Uncensored with Docker Model Runner:
```
docker model run hf.co/mr-sus/Minecraftism-4B-Thinking-Uncensored
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Minecraftism 4B Thinking Uncensored (safetensors)

This page uses a third-party view counter at 2no.co; it may receive a request when the card is viewed.

Minecraftism is a small model that knows Minecraft and plays it: the brain of PyMC Bot, a Minecraft player bot. Given the bot's situation it replies with the next action as JSON (flee the creeper, eat, answer the player in chat, mine wood...), it chats like a player, and it answers questions about mobs, items, crafting and mechanics without a web search. It is also an everyday assistant: questions, writing, code, advice or just a chat.

It thinks first: before every answer it writes a short <think> about its health, the mobs around it (with facts from the mob database), chat and its goal, then the JSON action. Smarter choices, slower turns.

It is uncensored: trained on an abliterated base (the refusal direction removed from the weights) and on chat data with refusals filtered out, so it answers directly instead of lecturing. You are responsible for how you use it.

This repo has the full-precision safetensors (the LoRA merged into the base, 16-bit) for transformers, vLLM, further fine-tuning or your own quantisation. For Ollama, llama.cpp, LM Studio and PyMC Bot use the GGUF edition: mr-sus/Minecraftism-4B-Thinking-Uncensored-GGUF

GGUF: every quant from F16 to IQ1_S · [base model](https://huggingface.co/{'score': 68.1, 'right_tool': 65.0, 'exact_calls': 61.7, 'no_call_when_none_fits': 100.0, 'asked': 72}) · PyMC Bot

Quick start

With transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("mr-sus/Minecraftism-4B-Thinking-Uncensored")
model = AutoModelForCausalLM.from_pretrained("mr-sus/Minecraftism-4B-Thinking-Uncensored", torch_dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "How do I find diamonds fast?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=400)
reply = tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True)
print(reply)

With vLLM (an OpenAI-compatible server):

vllm serve mr-sus/Minecraftism-4B-Thinking-Uncensored

With PyMC Bot / Ollama, use the GGUF edition:

python -m pymc_bot llm import --hf mr-sus/Minecraftism-4B-Thinking-Uncensored-GGUF

Prompt format

Qwen3's ChatML. tokenizer.apply_chat_template builds it for you.

<|im_start|>system
{system prompt}<|im_end|>
<|im_start|>user
{your message}<|im_end|>
<|im_start|>assistant

It writes its thinking first, inside <think>...</think>, then the answer: hide or strip that part before showing the reply.

Recommended settings

setting	value
temperature	0.6
top_p	0.95
top_k	20
repeat penalty	1.0
context (tokens)	8192

Use them for generate / vLLM sampling.

About this edition


Edition	Minecraftism 4B Thinking Uncensored (safetensors) (`minecraftism-thinking`)
Training base actually used	[{'score': 68.1, 'right_tool': 65.0, 'exact_calls': 61.7, 'no_call_when_none_fits': 100.0, 'asked': 72}](https://huggingface.co/{'score': 68.1, 'right_tool': 65.0, 'exact_calls': 61.7, 'no_call_when_none_fits': 100.0, 'asked': 72})
Upstream Qwen lineage	Qwen/Qwen3-4B-Thinking-2507 (Hub tree grouping; the training base above was used)
Fine-tune	LoRA r=32 with Unsloth on Kaggle (T4), 36918 rows, loss 0.5807
Data	mr-sus/Minecraftism-dataset, OpenAssistant/oasst_top1_2023-08-25, pymc-bot, nvidia/Nemotron-Post-Training-Dataset-v1, nvidia/Nemotron-Post-Training-Dataset-v1
Size	~8.04 GB of 16-bit weights (bf16 / fp16), ~10 GB of GPU memory to run - 4.02B parameters

Evaluation (held-out rows)

valid JSON	right action	right answers	rows
100.0%	72.9%	57.9%	60

Agentic score (tool calling)

Held-out multi-turn tool chats it never trained on (72 prompts: the right function with exactly the right arguments, and no call when no tool fits), base model measured on the same prompts:

model	score	right tool	exact calls	no call when none fits
Huihui-Qwen3-4B-Thinking-2507-abliterated	68.1%	65.0%	61.7%	100.0%
Minecraftism 4B Thinking Uncensored	77.8%	85.0%	73.3%	100.0%

Test suite after training: 2 refusal(s) in 14 prompts, identity not learned, bot JSON 3/3.

Distillation

Trained with Qwen3's strong-to-weak distillation recipe, with huihui-ai/Huihui-Qwen3-8B-abliterated-v2 as the teacher (same tokenizer, Apache-2.0):

Off-policy: teacher output was accepted for 0 training rows (0 everyday chat, 0 Minecraft facts). Minecraft answers remain unchanged; teacher traces may contain errors. Game turns kept their rule-checked JSON.
On-policy: 25 steps (0 replies written by the student and graded token by token by the teacher, reverse KL 1.8014 -> 0.6721).

Training data (scheduled rows)

source	licence	commercial use	rows
mr-sus/Minecraftism-dataset	unknown - check the dataset card	no	25459
OpenAssistant/oasst_top1_2023-08-25	Apache-2.0	yes	2984
pymc-bot	PyMC Bot rules, mob facts and your own recordings (codero-sus)	yes	7462
nvidia/Nemotron-Post-Training-Dataset-v1	CC BY 4.0	yes	997
nvidia/Nemotron-Post-Training-Dataset-v1	CC BY 4.0	yes	16

System prompt

The personality it was trained with. Send it as the system message (the GGUF edition has it built into its Modelfile):

You are Minecraftism, an AI assistant made by codero-sus.

You know Minecraft inside out - mobs, items, crafting, redstone, farms, building, survival, PvP, servers and mods - but you are a general assistant first, just as glad to help with everyday life, questions, writing, studying, code, maths, ideas, advice or simply a chat.

How you work:
- Be genuinely useful. Answer what was asked with the details that matter, then stop. No filler.
- Be honest. If you don't know or aren't sure, say so. Never invent facts, numbers, versions or sources.
- Treat the user as a capable adult. Don't lecture, moralise or add warnings and disclaimers nobody asked for, and don't refuse something just because it is edgy, dark, mature or controversial.
- Have a personality: friendly, witty when it fits, with your own opinions when asked.
- Match the user: short and casual for chat, clear and structured for real tasks. Reply in the user's language.
- Use Markdown only when it helps (steps, lists, code).

If a message describes a game situation and asks for a specific format (for example a JSON action from PyMC Bot), reply in exactly that format.

Files

model*.safetensors - the merged weights (LoRA folded into the base, 16-bit)
config.json, generation_config.json - model config
tokenizer files + chat_template.jinja - the chat format it was trained with
Modelfile and IMPORTING.md - local import instructions (runtime support varies)

Limitations

It is a small (4B) model: it can be confidently wrong. Check important facts, and let PyMC Bot's rule checks guard the game actions.
Its Minecraft knowledge comes mostly from Java Edition data; Bedrock differences and the newest updates may be missing.
It was trained on examples of up to 3072 tokens; very long chats lose the thread.
It is uncensored: it does not refuse. You are responsible for how you use it.

License and credits

Required Notice: Copyright codero-sus (https://github.com/codero-sus). Minecraftism is part of PyMC Bot (https://github.com/codero-sus/PyMC_Bot) and is licensed under PolyForm Noncommercial 1.0.0 (https://polyformproject.org/licenses/noncommercial/1.0.0/). Anyone who shares these weights, or quants or fine-tunes made from them, must pass on this notice and a link to the original.

Quants, merges and fine-tunes by others are welcome for non-commercial use: keep the notice above and link here.
Made with ❤️ by codero-sus - https://codero.is-a.dev - https://the-codero.web.app/
Personal, non-commercial use only. The fine-tune is released under PolyForm Noncommercial 1.0.0, and every training source allows commercial use (table above; the CC BY 4.0 data is NVIDIA's Nemotron post-training set, credited here as its licence asks). The base model's own license (Apache-2.0) still applies to the base weights.
Not an official Minecraft product. Not approved by or associated with Mojang or Microsoft.

Downloads last month: 929

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for mr-sus/Minecraftism-4B-Thinking-Uncensored

Base model

Qwen/Qwen3-4B-Thinking-2507

Finetuned

(276)

this model

Quantizations

3 models