The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models

Human feedback is central to the alignment of Large Language Models (LLMs).However, open questions remain about methods (how), domains (where), people(who) and objectives (to what end) of feedback processes. To navigate thesequestions, we introduce PRISM, a dataset that maps the sociodemographics andstated preferences of 1,500 diverse participants from 75 countries, to theircontextual preferences and fine-grained feedback in 8,011 live conversationswith 21 LLMs. With PRISM, we contribute (i) wider geographic and demographicparticipation in feedback; (ii) census-representative samples for two countries(UK, US); and (iii) individualised ratings that link to detailed participantprofiles, permitting personalisation and attribution of sample artefacts. Wetarget subjective and multicultural perspectives on value-laden andcontroversial issues, where we expect interpersonal and cross-culturaldisagreement. We use PRISM in three case studies to demonstrate the need forcareful consideration of which humans provide what alignment data.

Further reading#

• Access Paper in arXiv.org