OpenAssistant Conversations (OASST)
DatasetFree161K human-written messages in 35 languages with quality ratings.
Capabilities8 decomposed
multi-turn conversation tree dataset with branching preference paths
Medium confidenceProvides 66,497 conversation trees with 161,443 messages where each conversation branches into multiple continuations, enabling models to learn from human preference comparisons between different response paths. The branching structure is stored as a directed acyclic graph (DAG) where each message node can have multiple child responses, allowing RLHF algorithms to compare preferred vs non-preferred continuations at scale without requiring explicit pairwise annotations.
Implements explicit conversation branching as DAG structures rather than flat turn sequences, enabling direct preference comparison between alternative continuations without synthetic pair generation. The tree structure preserves the full context path for each response, allowing models to learn from natural human preference divergence points.
Unlike flat instruction datasets (Alpaca, ShareGPT) or synthetic preference pairs, OASST's branching structure captures real human preference diversity at scale with 161K messages from 13K+ annotators, making it significantly more robust for RLHF than datasets with single-path conversations.
human quality ratings and comparative ranking annotations
Medium confidenceEach message in the dataset includes human-assigned quality ratings (typically on a 1-5 scale) and comparative rankings where annotators explicitly ranked multiple responses to the same prompt. These ratings are aggregated across multiple annotators per message, providing consensus quality scores that can be used as reward signal targets or for filtering low-quality training data. The multi-annotator approach reduces individual bias and provides confidence estimates via inter-rater agreement metrics.
Implements multi-annotator consensus scoring where each message is rated by multiple independent human raters, with explicit comparative ranking annotations between responses. This approach provides both absolute quality scores and relative preference signals in a single dataset, enabling both regression-based and ranking-based reward model training.
Compared to single-annotator datasets or synthetic preference pairs, OASST's multi-rater approach provides statistically grounded quality signals with measurable inter-rater agreement, making it more reliable for training robust reward models than datasets with single judgments per example.
multilingual conversation dataset with 35-language coverage
Medium confidenceContains 161,443 messages across 35 languages including low-resource languages, collected through a distributed volunteer annotation process. Each conversation is tagged with its primary language, and the dataset includes both monolingual conversations and code-switching examples. The language distribution is uneven (English-heavy) but provides genuine human-written content in non-English languages rather than machine translations, enabling training of multilingual instruction-following models.
Provides genuinely human-written multilingual conversations from native speakers rather than machine-translated English content, with explicit language tagging and support for code-switching. The volunteer-driven collection process ensures natural language use patterns specific to each language community.
Unlike machine-translated instruction datasets or English-only collections, OASST captures authentic multilingual instruction-following patterns from 13K+ native speakers across 35 languages, providing significantly more natural and culturally appropriate training data for non-English models.
toxicity and safety annotations with label taxonomy
Medium confidenceMessages are annotated with toxicity labels and safety-relevant metadata using a structured taxonomy that includes categories like hate speech, violence, sexual content, and other harmful content types. Annotations are provided by human raters trained on the taxonomy, with multiple raters per message to establish consensus. The dataset includes both binary toxicity flags and fine-grained category labels, enabling training of content moderation models and safety-aware RLHF.
Implements structured toxicity taxonomy with multi-category fine-grained labels (hate speech, violence, sexual content, etc.) rather than binary toxicity flags, enabling nuanced safety analysis and category-specific moderation. Multi-annotator consensus approach provides confidence estimates for ambiguous cases.
Compared to single-label toxicity datasets or synthetic safety annotations, OASST provides human-validated multi-category toxicity labels from multiple raters on real conversational data, enabling more sophisticated safety-aware training than binary filtering approaches.
instruction-response pair extraction with context preservation
Medium confidenceThe dataset can be processed to extract instruction-response pairs while preserving full conversation context, enabling both single-turn instruction tuning and multi-turn dialogue training. The extraction process maintains parent-child relationships in the conversation tree, allowing models to learn from the full dialogue history leading up to each response. This differs from flat instruction datasets by preserving the sequential dependency structure and enabling context-aware response generation.
Enables extraction of instruction-response pairs while preserving full conversation context and parent-child relationships from the tree structure, rather than flattening to isolated pairs. This allows training models that understand dialogue history and can generate context-aware responses.
Unlike flat instruction datasets (Alpaca, Self-Instruct) that provide isolated instruction-response pairs, OASST's tree structure enables extraction of context-aware training examples where the model learns from full conversation history, producing more natural multi-turn dialogue behavior.
volunteer contributor metadata and annotation provenance tracking
Medium confidenceThe dataset includes metadata about the 13,000+ volunteer annotators who contributed messages and ratings, including their language preferences, annotation history, and quality metrics. This enables analysis of annotator bias, identification of high-quality contributors, and filtering of data based on annotator reliability. Provenance tracking allows researchers to understand which annotators contributed which messages and ratings, enabling weighted training schemes that prioritize high-quality annotators.
Provides explicit annotator IDs and contribution tracking across 13K+ volunteers, enabling analysis of annotator-level bias and reliability rather than treating all annotations as equally trustworthy. This enables weighted training schemes that account for annotator quality variation.
Unlike datasets with anonymous or aggregated annotations, OASST's annotator provenance tracking enables identification of high-quality contributors and implementation of annotator-weighted training, improving robustness against individual annotator bias.
conversation metadata and contextual filtering
Medium confidenceEach conversation includes metadata such as conversation ID, creation timestamp, language, and conversation-level quality assessments. This enables filtering and stratification of the dataset by temporal patterns, language, or quality tier. The metadata structure allows researchers to create balanced training splits that control for language distribution, conversation quality, or temporal effects, and to analyze how conversation-level properties correlate with response quality.
Provides conversation-level metadata enabling stratified sampling and filtering by language, quality, and temporal patterns, rather than treating all conversations as interchangeable. This allows controlled experiments that account for dataset composition effects.
Compared to datasets without conversation-level metadata, OASST enables stratified train/val/test splits that control for language distribution and quality variation, reducing confounding factors in model evaluation.
open-source dataset distribution with huggingface integration
Medium confidenceThe dataset is published on HuggingFace Datasets Hub with standardized loading APIs, version control, and documentation. This enables one-line dataset loading via the HuggingFace datasets library, automatic caching, and integration with popular ML frameworks (PyTorch, TensorFlow). The open-source distribution includes data cards documenting dataset composition, limitations, and intended use, facilitating reproducible research and transparent dataset governance.
Provides standardized HuggingFace Datasets Hub integration with one-line loading, automatic caching, and version control, rather than requiring manual download and parsing. Includes comprehensive data cards documenting composition, limitations, and ethical considerations.
Compared to datasets distributed as raw files or custom APIs, OASST's HuggingFace integration enables seamless integration with popular ML frameworks, automatic caching, and transparent dataset governance through standardized documentation.
Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.
Related Artifactssharing capabilities
Artifacts that share capabilities with OpenAssistant Conversations (OASST), ranked by overlap. Discovered automatically through the match graph.
UltraChat 200K
200K high-quality multi-turn dialogues for instruction tuning.
Capybara
Multi-turn conversation dataset for steerable models.
WildChat
1M+ real user-AI conversations with demographic metadata.
Nectar
183K multi-turn preference comparisons for alignment.
MT-Bench
Multi-turn conversation benchmark — 80 questions, 8 categories, GPT-4 as judge.
prompt-optimizer
An AI prompt optimizer for writing better prompts and getting better AI results.
Best For
- ✓ML researchers implementing RLHF pipelines for dialogue models
- ✓Teams training instruction-following models with preference learning
- ✓Organizations building reward models for LLM alignment
- ✓Researchers training reward models with human preference signals
- ✓Teams implementing quality-aware curriculum learning for dialogue models
- ✓Organizations conducting human evaluation studies on LLM outputs
- ✓Teams building multilingual LLM assistants
- ✓Researchers studying cross-lingual instruction following
Known Limitations
- ⚠Branching structure requires custom parsing logic — no standardized tree traversal API provided
- ⚠Not all conversation paths are equally balanced; some branches have significantly fewer continuations than others
- ⚠Preference annotations are implicit (via quality ratings) rather than explicit pairwise comparisons, requiring post-processing to extract training pairs
- ⚠Quality ratings are subjective and may not align with downstream task performance metrics
- ⚠Inter-rater agreement varies significantly across message types; some categories have κ < 0.4
- ⚠Ratings are sparse for some messages (only 1-2 annotators) reducing confidence in consensus scores
Requirements
Input / Output
UnfragileRank
UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.
About
Human-generated conversational dataset created by over 13,000 volunteers through the Open Assistant project. Contains 161,443 messages across 66,497 conversation trees in 35 languages. Each message has human quality ratings, labels, and toxicity annotations. Multi-turn conversations with branching paths allow preference learning. The largest human-written (not LLM-generated) instruction dataset available. Used to train OpenAssistant models and widely adopted for RLHF research.
Categories
Alternatives to OpenAssistant Conversations (OASST)
The GitHub for AI — 500K+ models, datasets, Spaces, Inference API, hub for open-source AI.
Compare →FLUX, Stable Diffusion, SDXL, SD3, LoRA, Fine Tuning, DreamBooth, Training, Automatic1111, Forge WebUI, SwarmUI, DeepFake, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, RunPod, Kaggle, NoteBooks, ControlNet, TTS, Voice Cloning, AI, AI News, ML, ML News,
Compare →Are you the builder of OpenAssistant Conversations (OASST)?
Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.
Get the weekly brief
New tools, rising stars, and what's actually worth your time. No spam.
Data Sources
Looking for something else?
Search →