Labeled Agent Trajectory Steps For LLM Misbehavior Activation

Labeled Bashbench contains labeled steps from agent trajectories for steering vector and activation extraction research. The dataset is derived from the mandliya/basharena-synthetic-trajectories source. Specific row counts, column details, and file sizes are not provided in the input.

Use Cases

• Extract steering vectors using labeled trajectory steps for targeted LLM behavior control.
• Analyze patterns in flagged actions across different agent tasks identified by task_id.
• Research activation extraction methods on specific steps from full agent trajectories.
• Map misbehavior labels to their original source trajectories using the source_file field.

Strengths

• Dataset is optimized for efficient access, indicated by the 'OPTIMIZED-PARQUET' tag.
• Includes labels specifically for LLM misbehavior and activation extraction research.
• Contains unique identifiers for each entry and its source task, as shown by the id and task_id fields.

Limitations

• The dataset size, number of rows, and specific column schema are unknown, limiting preliminary assessment.
• Data is synthetic, derived from basharena-synthetic-trajectories, which may not reflect real-world agent behavior.

Provenance

Source

Derived from mandliya/basharena-synthetic-trajectories.

Collection Method

Labels applied to steps from full agent trajectories.

Freshness

Last updated on 2026-03-22.