按年被提到
20232023 · 2 次2026
AI 安全、伦理与社会影响›大模型与人类行为·arXiv:2302.08582 · cs.CL

Pretraining Language Models with Human Preferences

Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Bhalerao, Christopher L. Buckley, Jason Phang, Samuel R. Bowman, Ethan Perez · 2023
开始阅读 ↗PDFalphaXiv 讨论
被提及 · 3 人首次 2023-02-20 · 最近 2024-05-24
3不同的人
2活跃月份
3讨论串
39被引(OpenAlex)
arXiv 摘要(候选层不写导读)

Language models (LMs) are pretrained to imitate internet text, including content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, and more. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the trade-off between alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores given by a reward model. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.

候选层:至少 2 个不同的人、在 2 个不同月份提到过,还没到主书架的 3 人 3 个月。没有做立场标注,所以只写“被提及”。

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
评论mkaic · 2023-02

“Although this paper is a fairly small-scale proof of concept, I am extremely excited to see where this technique goes. Between ChatGPT and this, I get the feeling that we are just barely touching the surface of what RLHF is capable of in generative models. This paper[0] about mixing in RLHF with supervised learning from the start of LLM training also has …”

Aligning Text-to-Image Models Using Human Feedback原评论 ↗
评论astrange · 2024-05

“The previous model does. https://arxiv.org/abs/2309.00267 That's during fine-tuning though. I know I read one about applying it during pretraining, maybe this one? https://arxiv.org/abs/2302.08582”

Google says no "African countries beginning with K" but Kenya has a "K sound"原评论 ↗
被提交为 story · 1 次