DefCPM-2B aims to be a lightweight truth-seeking sycophancy-free LLM assistant.
After experimenting with several LLMs (some of which I have shared on this account), I came to the conclusion that many Instruction-Tuned models would rather sacrifice accuracy that to dissatisfy the user (a phenomenon known as sycophancy). Therefore, I created this dataset to tackle with this exact problem, to encourage LLMs to give precise answers, teaching them to directly contest false user premises.
This model was fine-tuned (3 epochs, 0,00001 learning rate) with said dataset as an application of the concept. It adopted the expected behavior surprisingly well.
I generally recommend a Q4_K_M or an IQ4_XS quantization for deployment on edge devices.
Enjoy!
DefCPM stands for 'Definitive CPM'.