HALO Gemini 3 Flash AppWorld: Agent Execution Traces on Test-Normal Split

168 agent execution traces of the Gemini 3 Flash model interacting with AppWorld's simulated app ecosystem. The dataset, created by inference-net, captures full span-level execution detail from the test-normal benchmark split. It was last updated on Hugging Face on April 29, 2026.

Use Cases

• Analyzing agent reasoning and decision-making patterns based on the captured execution traces.
• Benchmarking large language model performance on interactive, tool-use tasks within a simulated ecosystem.
• Training or fine-tuning agentic models using detailed examples of successful and failed interactions.
• Studying the failure modes and error propagation of AI agents in multi-step application workflows.

Strengths

• Contains 168 distinct execution traces, providing a substantive sample for analysis.
• Captures full span-level detail of the agent's interaction, likely offering granular step-by-step data.
• Focuses on a specific, controlled benchmark (AppWorld) and split (test-normal) for consistent evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual spans is unknown, which may limit suitability assessment for certain tasks.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

huggingface

Collection Method

Agent execution traces generated from running the Gemini 3 Flash model on the AppWorld benchmark.

Freshness

Last updated 2026-04-29 07:16:11; freshness should be verified.