Harbor Adapter: Agent Execution Trajectories from Large-Scale Evaluation

Agent execution trajectories from a large-scale agentic evaluation. Each trajectory captures a single (agent, model) attempt at a task, including step logs, tool calls, model outputs, and verifier scoring. The dataset was authored by kendx and last updated on 2026-06-08.

Use Cases

• Benchmarking agent performance across different models based on verifier scoring.
• Analyzing agent decision-making patterns based on full step logs.
• Studying tool usage and environment interaction patterns based on tool calls and environment returns.
• Training or fine-tuning agent models based on execution trajectories.

Strengths

• Captures full execution trajectories end-to-end for each (agent, model, task) combination.
• Includes verifier scoring for each trajectory, providing a performance metric.
• Designed for filtering to specific slices of interest before pulling data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Collected from a large-scale agentic evaluation.

Freshness

Last updated 2026-06-08 01:19:29; freshness should be verified.