107 lines
4.0 KiB
Markdown
107 lines
4.0 KiB
Markdown
---
|
|
license: gpl-3.0
|
|
language:
|
|
- es
|
|
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
|
tags:
|
|
- dpo
|
|
- ai-detection
|
|
- paraphrase
|
|
library_name: transformers
|
|
datasets:
|
|
- pymlex/ai-generated-texts
|
|
metrics:
|
|
- accuracy
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Qwen2.5-0.5B-Human
|
|
|
|
DPO fine-tune of [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) that paraphrases Spanish academic abstracts to reduce AI-detection scores from [danibor/oculus-v2.0-multilingual](https://huggingface.co/danibor/oculus-v2.0-multilingual).
|
|
|
|
## Overview
|
|
|
|
Preference pairs for optimisation come from [pymlex/ai-generated-texts](https://huggingface.co/datasets/pymlex/ai-generated-texts). The corpus is [Flaglab/academic-knowledge-abstracts-es](https://huggingface.co/datasets/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 | MCC | ROC-AUC | F1 |
|
|
| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
|
|
| base | validation | 1107 | 0.6550 | 1.4619 | 0.6712 | 0.0000 | n/a | 0.8032 |
|
|
| base | test | 1112 | 0.6532 | 1.5581 | 0.6655 | 0.0000 | n/a | 0.7991 |
|
|
| fine-tuned | validation | 1107 | 0.2264 | -2.0100 | 0.1716 | 0.0000 | n/a | 0.2930 |
|
|
| fine-tuned | test | 1112 | 0.2391 | -1.8733 | 0.1835 | 0.0000 | 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](https://github.com/pymlex/ai-text-detector-tricking).
|
|
|
|
## Citation
|
|
|
|
If you found this model useful, please cite it as:
|
|
|
|
```bibtex
|
|
@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}
|
|
}
|
|
```
|
|
|
|
```bibtex
|
|
@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}
|
|
}
|
|
```
|
|
|
|
```bibtex
|
|
@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}
|
|
}
|
|
```
|
|
|
|
```bibtex
|
|
@misc{oculus2026,
|
|
title = {{Oculus 2.0 Multilingual AI Text Detector}},
|
|
author = {danibor},
|
|
year = {2026},
|
|
url = {https://huggingface.co/danibor/oculus-v2.0-multilingual}
|
|
}
|
|
```
|
|
|
|
```bibtex
|
|
@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. |