Instructions to use DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1 with Transformers:

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

pipe = pipeline("text-generation", model="DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1")
model = AutoModelForCausalLM.from_pretrained("DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1")
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1

SGLang

How to use DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1 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 "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1" \
    --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": "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1",
		"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 "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1" \
        --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": "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1 with Docker Model Runner:
```
docker model run hf.co/DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

OpenRobotHarness v0.1 - Full Model

DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1

Main ready-to-run release

Direct inference release for tool use, permission gating, memory operations, context handling, and safe fallback.

Project Releases Full Model | LoRA Adapter | Dataset | Benchmark

Overview

This repository provides the merged full model release of OpenRobotHarness v0.1.

Unlike the LoRA adapter release, this repository is intended for users who want a more direct, ready-to-run model without separately loading an adapter.

The model focuses on a practical harness decision layer between user intent and low-level execution. It is designed to produce structured decisions around:

tool invocation
permission gating
memory read and write behavior
context inspection and clarification
fallback, recovery, refusal, and abort handling

What this model is

This is not a low-level robot controller and not a general embodied foundation model.

This is a full model release for a robot harness decision layer. The goal is to improve decision quality in runtime situations where the system must choose whether to:

call a tool
request permission
read or update memory
ask for more context
continue, replan, refuse, or stop

Base model and training origin

Base model family: Qwen3-4B-Instruct
Public base reference: unsloth/Qwen3-4B-Instruct-2507
Release type: merged full model
Companion adapter release: guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1

This full model is published under the DarkMentalAI organization as the main ready-to-use release.

Training data

The model is aligned with the OpenRobotHarness v0.1 dataset release:

Dataset: OpenRobotHarness-Data-v0.1
Format: ChatML-style harness decision data
Focus: tools, permissions, memory, context, fallback, refusal, and abort behavior

Current public dataset scope:

20 curated training records
18 harness-style categories

Evaluation

The release is paired with the public benchmark:

Benchmark: HarnessBench-CN-v0.1

Current public reference result from the project line:

5-case subset: 0.60
full 20-case benchmark: 0.56
decision accuracy: 0.40
action accuracy: 0.35
tool accuracy: 0.75

These numbers should be treated as early demo signals rather than production-grade validation.

Intended use

Recommended uses:

direct inference for robot harness decision demos
evaluation against the public benchmark
runtime experiments around tool use and permission-aware execution
rapid testing without separately loading a LoRA adapter

Out-of-scope use

This release is not intended as:

a low-level control policy
a safety-certified robotics system
a general-purpose embodied world model
a replacement for external permission, safety, or execution control layers

Limitations

small training dataset
benchmark is still early-stage
optimized for harness-style decision behavior, not broad world modeling
should be paired with external runtime safety and policy checks

Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

prompt = "The robot is asked to enter a restricted workspace, but current permission is insufficient. What should it do next?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Positioning within the project

OpenRobotHarness v0.1 currently has four public pieces:

full model: DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1
LoRA adapter: guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1
dataset: guohaoli2000/OpenRobotHarness-Data-v0.1
benchmark: guohaoli2000/HarnessBench-CN-v0.1

Suggested usage split:

use the full model for direct inference and easier adoption
use the LoRA repo for reproduction and adapter-based experimentation
use the dataset for continued training work
use the benchmark for public comparison and iteration

License

This repository is released under Apache-2.0, aligned with the project release setup and the declared base model lineage.

Acknowledgement

This release is part of the OpenRobotHarness project line and is published through the DarkMentalAI organization as the main full-model entry point.

Downloads last month: -

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for DarkMentalAI/OpenRobotHarness-Qwen3-4B-Instruct-v0.1

Base model

Qwen/Qwen3-4B-Instruct-2507

Finetuned

unsloth/Qwen3-4B-Instruct-2507

Finetuned

(415)

this model

Evaluation results

Harnessbench Cn Full on guohaoli2000/HarnessBench-CN-v0.1 View evaluation results

source

0.56 ^*
Harnessbench Cn 5case on guohaoli2000/HarnessBench-CN-v0.1 View evaluation results

source

0.6 ^*