tlc4418/1.4b-policy_preference_data_gold_labelled

Preference dataset using labels from the [AlpacaFarm dataset](https://huggingface.co/datasets/tatsu-lab/alpaca_farm), generated answers from a [1.4b fine-tuned Pythia policy model](https://huggingface.co/tlc4418/pythia_1.4b_sft_policy), and labelled using the AlpacaFarm '[reward-model-human](https://github.com/tatsu-lab/alpaca_farm#downloading-pre-tuned-alpacafarm-models)' as a gold reward model.

Used to train reward models in '[Reward Model Ensembles Mitigate Overoptimization](https://arxiv.org/abs/2310.02743)'

原始信息汇总

数据集概述

数据来源

• 使用AlpacaFarm数据集的标签。
• 生成答案来自1.4b微调的Pythia策略模型。
• 使用AlpacaFarm的reward-model-human作为黄金奖励模型进行标注。

应用场景

• 用于训练奖励模型，相关研究为《Reward Model Ensembles Mitigate Overoptimization》。