HelpSteer is an open-source Helpfulness Dataset (CC-BY-4.0) that supports aligning models to become more helpful, factually correct and coherent, while being adjustable in terms of the complexity and verbosity of its responses.
Leveraging this dataset and SteerLM, we train a Llama 2 70B to reach 7.54 on MT Bench, the highest among models trained on open-source datasets based on MT Bench Leaderboard as of 15 Nov 2023.
This model is available on… See the full description on the dataset page:
https://huggingface.co/datasets/nvidia/HelpSteer.