OpenAssistant Conversations 2 (OASST2)
Human-written assistant conversations in 35 languages with quality rankings, incl. Russian.
Overview
OpenAssistant Conversations Release 2 (oasst2) is a fully human-generated assistant dialogue corpus built by more than 13,000 volunteers of the OpenAssistant project. Contributors wrote both user prompts and assistant replies, then graded each other's messages for quality, helpfulness and safety — producing ranked conversation trees rather than flat pairs.
Contents
135,174 messages organised into 13,854 conversation trees across 35 languages; English, Spanish, Russian, German and Chinese dominate the distribution. Every node carries its role, language, human quality labels, rank among sibling replies and toxicity scores, so one dump serves SFT, reward modelling and preference tuning alike.
Why it matters for neuronet training & evaluation
Apache-2.0 with explicit commercial use, no synthetic-data clauses and strong Russian coverage make oasst2 a safe default for instruction-tuning ru/en neuronets. Sibling rankings convert directly into DPO pairs, and held-out trees double as a chat-quality evaluation set.