Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.
核心贡献 · Key contributions
展示了通过人类反馈微调（RLHF）使语言模型与用户意图对齐，1.3B 参数的 InstructGPT 优于 175B 的 GPT-3。 Shows fine-tuning with human feedback (RLHF) aligns language models to user intent, with 1.3B InstructGPT preferred over 175B GPT-3.
InstructGPT 在 TruthfulQA 上提升了真实性，并在被要求尊重时减少了 25% 的有毒输出。 InstructGPT improves truthfulness on TruthfulQA and reduces toxic outputs by 25% when prompted respectfully.
RLHF 对齐成本相对于预训练较低；1.3B InstructGPT 以 100 倍少的参数优于 175B GPT-3。 RLHF alignment cost is modest relative to pretraining; 1.3B InstructGPT outperforms 175B GPT-3 with 100x fewer parameters.
混合预训练梯度（PPO-ptx）减轻了在 SQuAD 和 HellaSwag 等 NLP 基准上的对齐代价。 Mixing pretraining gradients (PPO-ptx) mitigates alignment tax on NLP benchmarks like SQuAD and HellaSwag.
InstructGPT 将指令遵循泛化到非英语和代码任务，尽管监督有限。 InstructGPT generalizes instruction-following to non-English and code tasks despite limited supervision.
保留的标注者以相似比率偏好 InstructGPT 而非 GPT-3，表明泛化到训练标注者之外。 Held-out labelers prefer InstructGPT over GPT-3 at similar rates, showing generalization beyond training labelers.
局限 · Limitations
InstructGPT 仍会产生有毒、有偏见或虚假的输出，尤其在要求有害内容时。 InstructGPT still generates toxic, biased, or false outputs, especially when prompted for harmful content.
对齐针对特定标注者偏好，而非普世人类价值观；标注者人口统计有限。 Alignment is to specific labeler preferences, not universal human values; labeler demographics are limited.
即使采用 PPO-ptx 缓解，某些 NLP 数据集（如 SQuAD、DROP）上的性能下降仍然存在。 Performance regressions on some NLP datasets (e.g., SQuAD, DROP) persist even with PPO-ptx mitigation.
模型在错误前提指令上失败，对简单问题过度含糊，且难以处理多重约束。 Model fails on false premise instructions, over-hedges simple questions, and struggles with multiple constraints.
RLHF 可能无法扩展到超级智能系统；该方法依赖人类评估，对复杂任务变得困难。 RLHF may not scale to superhuman systems; method relies on human evaluation which becomes difficult for complex tasks.
论文章节 · Sections（共 19）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related work
3.1 高层方法论3.1 High-level methodology
3.2 数据集3.2 Dataset
3.3 任务3.3 Tasks
3.4 人类数据收集3.4 Human data collection
3.5 模型3.5 Models
3.6 评估3.6 Evaluation
4 结果4 Results
4.1 API 分布上的结果4.1 Results on the API distribution
4.2 在公开 NLP 数据集上的结果4.2 Results on public NLP datasets
4.3 定性结果4.3 Qualitative results
5.1 对对齐研究的启示5.1 Implications for alignment research