Instructions to use pymlex/Qwen2.5-0.5B-Human with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use pymlex/Qwen2.5-0.5B-Human with Transformers:

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

pipe = pipeline("text-generation", model="pymlex/Qwen2.5-0.5B-Human")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("pymlex/Qwen2.5-0.5B-Human")
model = AutoModelForMultimodalLM.from_pretrained("pymlex/Qwen2.5-0.5B-Human")
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 pymlex/Qwen2.5-0.5B-Human with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "pymlex/Qwen2.5-0.5B-Human"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pymlex/Qwen2.5-0.5B-Human",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/pymlex/Qwen2.5-0.5B-Human

SGLang

How to use pymlex/Qwen2.5-0.5B-Human 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 "pymlex/Qwen2.5-0.5B-Human" \
    --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": "pymlex/Qwen2.5-0.5B-Human",
		"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 "pymlex/Qwen2.5-0.5B-Human" \
        --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": "pymlex/Qwen2.5-0.5B-Human",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use pymlex/Qwen2.5-0.5B-Human with Docker Model Runner:
```
docker model run hf.co/pymlex/Qwen2.5-0.5B-Human
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen2.5-0.5B-Human

DPO fine-tune of Qwen/Qwen2.5-0.5B-Instruct that paraphrases Spanish academic abstracts to reduce AI-detection scores from danibor/oculus-v2.0-multilingual.

Overview

Preference pairs for optimisation come from pymlex/ai-generated-texts. The corpus is Flaglab/academic-knowledge-abstracts-es. For each train abstract, two base-model paraphrases are ranked by Oculus logit. DPO with beta = 0.1 increases the relative log-probability of the lower-logit completion. Retained pairs: 6396 from 8891 train abstracts with absolute logit gap at least 1.

Evaluation setup

Hardware: NVIDIA RTX 5090, Ubuntu Jupyter, CUDA 13.0+, bf16 training and inference. Post-training evaluation generates one paraphrase per validation and test abstract with the base and fine-tuned models, scores each output with Oculus, and treats label 1 as AI-generated at threshold 0.5 on detector probability.

During DPO, mean validation AI probability on a 276-text subset moved from 0.6740 at step 0 to 0.2437 at the last monitor step (-0.4303).

Results

Model	Split	n	mean prob	mean logit	accuracy	ROC-AUC	F1
base	validation	1107	0.6550	1.4619	0.6712	n/a	0.8032
base	test	1112	0.6532	1.5581	0.6655	n/a	0.7991
fine-tuned	validation	1107	0.2264	-2.0100	0.1716	n/a	0.2930
fine-tuned	test	1112	0.2391	-1.8733	0.1835	n/a	0.3100

Lower mean probability and MCC near zero indicate weaker detector response on model paraphrases under the AI-positive labelling convention.

Source code

The full pipeline is published on GitHub.

Citation

If you found this model useful, please cite it as:

@misc{zyukov2026qwenhuman,
  title         = {{Qwen2.5-0.5B-Human: DPO fine-tune against Oculus detector}},
  author        = {Zyukov, Alex},
  year          = {2026},
  url           = {https://huggingface.co/pymlex/Qwen2.5-0.5B-Human}
}

@misc{zyukov2026aitexttricking,
  title         = {{DPO Fine-Tuning Against Multilingual AI Text Detectors}},
  author        = {Zyukov, Alex},
  year          = {2026},
  url           = {https://github.com/pymlex/ai-text-detector-tricking},
  publisher     = {GitHub},
  organization  = {pymlex}
}

@misc{nicks2024detectors,
  title         = {{Language Model Detectors Are Easily Optimized Against}},
  author        = {Nicks, Cameron and Chua, Jeremy and Liu, Stephen and others},
  year          = {2024},
  eprint        = {2406.07490},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2406.07490}
}

@misc{oculus2026,
  title         = {{Oculus 2.0 Multilingual AI Text Detector}},
  author        = {danibor},
  year          = {2026},
  url           = {https://huggingface.co/danibor/oculus-v2.0-multilingual}
}

@misc{flaglab2025abstracts,
  title         = {{Academic Knowledge Abstracts Spanish}},
  author        = {Flaglab},
  year          = {2025},
  url           = {https://huggingface.co/datasets/Flaglab/academic-knowledge-abstracts-es}
}

The project is under GPL-3.0 license.