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.
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.
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).
Lower mean probability and MCC near zero indicate weaker detector response on model paraphrases under the AI-positive labelling convention.
The full pipeline is published on
GitHub.
1@misc{zyukov2026qwenhuman,
2 title = {{Qwen2.5-0.5B-Human: DPO fine-tune against Oculus detector}},
3 author = {Zyukov, Alex},
4 year = {2026},
5 url = {https://huggingface.co/pymlex/Qwen2.5-0.5B-Human}
6}
1@misc{zyukov2026aitexttricking,
2 title = {{DPO Fine-Tuning Against Multilingual AI Text Detectors}},
3 author = {Zyukov, Alex},
4 year = {2026},
5 url = {https://github.com/pymlex/ai-text-detector-tricking},
6 publisher = {GitHub},
7 organization = {pymlex}
8}
1@misc{nicks2024detectors,
2 title = {{Language Model Detectors Are Easily Optimized Against}},
3 author = {Nicks, Cameron and Chua, Jeremy and Liu, Stephen and others},
4 year = {2024},
5 eprint = {2406.07490},
6 archivePrefix = {arXiv},
7 primaryClass = {cs.CL},
8 url = {https://arxiv.org/abs/2406.07490}
9}
1@misc{oculus2026,
2 title = {{Oculus 2.0 Multilingual AI Text Detector}},
3 author = {danibor},
4 year = {2026},
5 url = {https://huggingface.co/danibor/oculus-v2.0-multilingual}
6}
1@misc{flaglab2025abstracts,
2 title = {{Academic Knowledge Abstracts Spanish}},
3 author = {Flaglab},
4 year = {2025},
5 url = {https://huggingface.co/datasets/Flaglab/academic-knowledge-abstracts-es}
6}
The project is under GPL-3.0 license.
