SODA: Million-Scale Social Dialogue Dataset from AI-Distilled Knowledge

SODA is a million-scale dialogue dataset covering a wide range of social interactions. Dialogues were generated by contextualizing social commonsense knowledge from the Atomic10x knowledge graph using a large language model (InstructGPT). The dataset was created by AllenAI and released on Hugging Face, with a last update recorded on 2023-01-04.

Use Cases

• Training social dialogue agents based on the wide range of social interactions described.
• Benchmarking the consistency and naturalness of generated conversations against the human-evaluated SODA dialogues.
• Studying the application of knowledge graphs (Atomic10x) for grounding language model outputs in social commonsense.

Strengths

• Million-scale size, which is described as the first publicly available dataset of this scale for social dialogue.
• Human evaluation indicates dialogues are more consistent, specific, and natural than prior human-authored benchmarks.
• Created by distilling knowledge from a structured knowledge graph (Atomic10x) using a state-of-the-art language model (InstructGPT).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last updated 2023-01-04; freshness should be verified for contemporary applications.
• The dataset's method of generation from a language model may introduce biases inherent to the model and source knowledge graph.

Provenance

Source

AllenAI, via Hugging Face.

Collection Method

Dialogues were distilled from InstructGPT by contextualizing social commonsense knowledge from the Atomic10x knowledge graph.

Freshness

Last updated 2023-01-04 09:24:32.