User Engagement Data For Grounded LLM Interest Evaluation

GISTBench contains 4.2 million content engagements from approximately 1,000 anonymized users for evaluating large language models. Created by Facebook, this benchmark tests an LLM's ability to infer user interests from interaction history and ground those inferences in specific evidence. The dataset was last updated in April 2026.

Use Cases

• Train models to predict a user's future content preferences based on their historical engagement features.
• Benchmark LLM performance on extracting and citing specific relevant items from a user's interaction history.
• Develop algorithms to detect and cluster distinct user interest profiles from engagement patterns.
• Evaluate the groundedness of LLM-generated user interest summaries against the evidence in the engagement records.

Strengths

• Contains over 4.2 million engagement records.
• Derived from a base of approximately 1,000 distinct users.

Limitations

• User base of ~1,000 may not represent broader population diversity.
• Specific column schema and data types are not publicly detailed in the input.

Provenance

Source

Facebook

Collection Method

Collected from anonymized user interactions with content such as videos, articles, and books.

Freshness

Last updated in April 2026.