This repository contains the 0.6B A-GRPO checkpoint released with AAPA: Adversarially Anchored Preference Alignment for Post-Training of Large Language Models.

AAPA is a plug-in framework that augments post-training objectives with a sentence-level adversarial anchoring signal. It compares policy rollouts with offline expert responses using a fixed lightweight discriminator, providing semantic grounding during preference optimization.

This checkpoint is trained from Qwen3-0.6B using the AAPA code release.

Resources

Paper: https://arxiv.org/abs/2509.25148
Hugging Face Paper: https://huggingface.co/papers/2509.25148
Code: https://github.com/IsFaqq/AAPA
Data: https://huggingface.co/datasets/Jingleqian/AAPA-data
8B checkpoint: https://huggingface.co/Jingleqian/AAPA-8B

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Jingleqian/AAPA-06B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Citation

bibtex

@article{aapa2025,
  title={AAPA: Adversarially Anchored Preference Alignment for Post-Training of Large Language Models},
  author={Faqiang Qian and Kang An and Weikun Zhang and Ziliang Wang and Xuhui Zheng and Liangjian Wen and Yong Dai and Mengya Gao and Yichao Wu},
  journal={arXiv preprint arXiv:2509.25148},
  year={2025}
}