OpenAssistant/oasst2

Dataset Overview

Dataset Name

• Name: Open Assistant Conversations Release 2 (OASST2)
• Alias: OASST2

Dataset Content

• Type: Dialogue dataset
• Structure: Contains message trees; each tree is rooted by an initial prompt message and may have multiple reply levels.
• Roles: Message roles are "assistant" and "prompter", alternating strictly throughout the dialogue.

Dataset Features

• Feature List:
  • message_id (string)
  • parent_id (string)
  • user_id (string)
  • created_date (string)
  • text (string)
  • role (string)
  • lang (string)
  • review_count (integer)
  • review_result (boolean)
  • deleted (boolean)
  • rank (integer)
  • synthetic (boolean)
  • model_name (string)
  • detoxify (struct with multiple toxicity scores, all float)
  • message_tree_id (string)
  • tree_state (string)
  • emojis (sequence of name and count, string and integer)
  • labels (sequence of name, value, count; string, float, integer)

Dataset Splits

• Training Set: 128,575 samples, size 158,850,455 bytes
• Validation Set: 6,599 samples, size 7,963,122 bytes

Dataset Size

• Download Size: 66,674,129 bytes
• Dataset Size: 166,813,577 bytes

Supported Languages

• Supports multiple languages, including but not limited to English, Spanish, Russian, German, Polish, Thai, etc.

Dataset Files

• Message Tree File: .trees.jsonl.gz
• Flattened Message File: .messages.jsonl.gz

Dataset Status

• Ready‑to‑Export: 13,854 trees, total 135,174 messages
• Full: 70,642 trees, total 208,584 messages

Additional Exports

• Garbage Messages: 19,296 matching messages
• Prompt Messages: 64,592 matching messages

Using HuggingFace Datasets

• Provides training and validation splits that can be directly loaded via HuggingFace Datasets.

Data Visualization

• Data visualized using Bunka technology, offering interactive map exploration of the content.