OpenAI TLDR: Human Feedback for Summarization Preference Modeling

UCL-DARK formatted human feedback data for training reward functions, as referenced in the paper from October 2023. The dataset is intended for sequence classification models within the Hugging Face ecosystem. Its specific scale and structure are not detailed in the provided metadata.

Use Cases

• Train sequence classification reward models based on human feedback data.
• Fine-tune language models for summarization preference prediction.
• Benchmark reinforcement learning algorithms for aligning text generation with human preferences.

Strengths

• Data is directly linked to a specific research paper (arXiv:2310.06452) from October 2023.
• Formatted specifically for compatibility with Hugging Face Transformers library.

Limitations

• Column names, row count, and file size are unknown, limiting suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The author recommends using a different dataset (openai/summarize_from_feedback) for new projects.

Provenance

Source

UCL-DARK, based on original data from OpenAI's summarize-from-feedback project.

Collection Method

Human feedback data, reformatted for sequence classification.

Freshness

Last updated 2023-10-26 09:52:20