LinuxArena Trajectories: Full Agent Evaluation Runs

14 model/policy combinations were evaluated across 10 environments, generating full agent trajectories. Each row contains one complete evaluation trajectory with every tool call, argument, output, error, and reasoning step. The dataset was created by eval-aware and last updated on March 28, 2026.

Use Cases

• Analyzing agent failure modes based on detailed error and output logs.
• Benchmarking model performance across different Linux environments.
• Training or fine-tuning agents using demonstrated reasoning and action sequences.
• Studying the relationship between agent policies and tool-call patterns.

Strengths

• Contains trajectories from 14 distinct model/policy combinations.
• Evaluations were conducted across 10 different environments.
• Provides two granularity levels of detail per action.

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 LinuxBench/LinuxArena evaluations.

Freshness

Last updated 2026-03-28 14:17:55; freshness should be verified.