Fine-tuned with Group Relative Policy Optimization (GRPO) to
rewrite AI-generated German academic text so that it passes GPTZero
detection while preserving semantic content.
Training details
Parameter
Value
Base model
Qwen/Qwen3-8B
Method
GRPO (TRL) + LoRA
Learning rate
5e-06
Batch size
2
Gradient accumulation
8
Max steps
50
Precision
bf16
Intended use
Academic text humanisation for German-language content. The model is
designed to be called via the HuggingFace Inference API from the
GhostWriter application.
