Evidence item

The InstructGPT paper documents reinforcement learning fine-tuned using human preferences, supporting a human-data component while not establishing the claim's broader metaphor about alienness. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 9741

Source: Training language models to follow instructions with human feedback · locator {} · excerpt hash sha256:0be4314f98bb33053e4fdb1efe0592ae0ca2db8781caca4f8e394a5e415973af

We focus on fine-tuning approaches to aligning language models. Specifically, we use reinforcement learning from human feedback (RLHF) ... This technique uses human preferences as a reward signal to fine-tune our models. We first hire a team of 40 contractors to label our data.