MiniCPM5-1B Humanizer Matches Human Writing on AI Detector
TL;DR. Researchers developed a 1B-parameter AI model, MiniCPM5-1B, using stacked SFT and DPO LoRAs, to match human writing on the RADAR AI detector. - The model runs locally on a Mac and achieves a P(AI) score of 0.37, identical to the human reference. - It significantly reduces AI boilerplate phrases from 0.6 per 1K tokens to 0.0, outperforming human text. - This lightweight solution avoids the need for larger, API-driven humanizers, democratizing access to human-like text generation.
- Stacked SFT + DPO LoRAs on MiniCPM5-1B achieve human-level P(AI) scores (0.37) on the RADAR AI detector.
- The 1B-parameter model and its adapters run locally on consumer hardware like a 24 GB Mac.
- The system eliminates AI-specific boilerplate phrases, even outperforming human text in this regard.
- The OptIQ 0.1.4 release provides tools for quantization, SFT/DPO training, and model serving.
Sources
- A 1B humanizer that matches human writing on an AI detector — mlx-optiq.com