Llama-3-Varco-Offsetbias-8B

Llama-3-Varco-Offsetbias-8B is NCSOFT's generative judge model that performs pairwise preference evaluation task. It is trained to be robust on various evaluation biases commonly found in evaluation models, such as length bias, empty reference bias, and familiar knowledge bias. The methodology for training the model is introduced in paper "OffsetBias: Leveraging Debiased Data for Tuning Evaluators". Specifically, Llama-3-Varco-Offsetbias-8B is built with Meta Llama-3-8B-Instruct. It is fine-tuned on evaluation datasets for judging various capabilities including instruction following, safety, and dialogue. The model is trained to perform pairwise preference evaluation, where Instruction, Output (a), Output (b) are given, and a better output to the instruction needs to be found.