Large Language Models
Sep 19, 2019
GPT-2 Model Enhanced with Human Feedback for Improved Task Performance
Sep 19, 2019
AI Summary
A 774M parameter GPT-2 language model has been fine-tuned using human feedback to better align with user preferences across various tasks. The model showed a tendency to copy sentences for summarization tasks, reflecting the labelers' preferences, which highlights the importance of human values in machine communication.
- The GPT-2 language model has been fine-tuned with human feedback to improve task performance.
- The fine-tuning process involved 60,000 human labels for summarization tasks and 5,000 for simpler tasks that required text continuation in various styles.
- The model's adjustments aimed to enhance safety techniques in machine communication, emphasizing the extraction of human values.