Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html> <h"... is not valid JSON

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

CompactionBench LME Text Subset

20 questions from LongMemEval-V2 filtered for text-answerable content.

Each task has ~160k tokens of web agent trajectory context (thoughts, actions, page states). All answers are confirmed present in the text context.

Used for testing context compaction in long-running agents.

Format

JSONL with one task per line. Each line is a CompactionBench TaskRow with:

  • task_id
  • context (~160k tokens)
  • question
  • gold_answer
  • metadata (domain, question_type, trajectories)

Source

Questions from LongMemEval-V2. Trajectories rendered from agent thought logs and action histories.

Downloads last month
23