OpenAssistant Conversations (OASST)

Extracts complete conversation trees from 66,497 human-authored dialogues where each message can have multiple child responses, creating a directed acyclic graph (DAG) structure. The dataset preserves branching paths where volunteers provided alternative continuations at decision points, enabling training on diverse response distributions for the same context. This tree structure is serializable to JSON with parent-child message IDs, allowing downstream systems to reconstruct full conversation histories or sample specific branches for preference learning.

Each message includes quality ratings from multiple human annotators (typically 3-5 raters per message) on dimensions like helpfulness, harmlessness, and honesty. The dataset provides aggregated scores (mean, median, or consensus) plus raw per-annotator ratings, enabling calculation of inter-rater reliability (Krippendorff's alpha, Fleiss' kappa) and identification of ambiguous examples. This multi-rater approach reduces individual bias and allows filtering by agreement threshold to create high-confidence training subsets.

Messages are annotated with toxicity scores and categorical safety labels (e.g., sexual content, violence, illegal activity, misinformation) applied by human annotators. The dataset exposes both binary flags (toxic/non-toxic) and continuous toxicity scores, plus detailed category breakdowns. This enables training safety classifiers, filtering harmful content, and analyzing the distribution of safety issues across conversation types and languages.

Contains 161,443 messages across 35 languages with uneven distribution (English-dominant but includes low-resource languages like Swahili, Vietnamese, Polish). The dataset structure allows filtering by language code and sampling balanced subsets across languages. This enables training multilingual models, analyzing language-specific conversation patterns, and studying how human preferences vary across linguistic and cultural contexts.

Automatically generates preference training pairs by comparing sibling responses (multiple continuations of the same prompt) using aggregated human quality ratings. For each prompt with N child responses, the system creates preference triplets (prompt, higher-rated_response, lower-rated_response) by ranking children by quality score. This avoids synthetic preference generation and grounds preference learning in actual human judgments, enabling direct training of reward models and DPO-style algorithms.

Flattens conversation trees into instruction-response pairs by treating each user message as an instruction and the following assistant message as the response. Handles multi-turn context by optionally including conversation history or using only the immediate prompt-response pair. This enables straightforward supervised fine-tuning (SFT) of language models without requiring preference learning infrastructure, suitable for baseline model training or quick prototyping.

Each conversation includes metadata tags or inferred categories (e.g., creative writing, coding, Q&A, general knowledge) enabling domain-specific filtering and analysis. While not explicitly documented as structured tags in the original dataset, the message content and conversation structure allow downstream systems to classify conversations by type. This enables creating domain-specific training subsets, analyzing model performance across task types, and studying how human preferences vary by domain.

161,443 messages collected from 13,000+ volunteer annotators through a crowdsourced platform (Open Assistant project), not generated by LLMs or synthetic methods. The annotation pipeline includes message creation, quality rating, toxicity labeling, and ranking by multiple independent raters. This human-centric approach ensures authentic conversational patterns, diverse writing styles, and genuine human preferences, though with inherent quality variance across annotators.