Horizon Tasks: 786 Agent Trajectories for Web, OS, DB, and Embodied Domains

786 task execution records across four domains: web, operating system, database, and embodied agent tasks. The data includes agent trajectories and plans generated by language models GPT-5-mini and Claude-4-Sonnet. Author xwang2775 uploaded this collection to Hugging Face, with a last recorded update in April 2026.

Use Cases

• Benchmarking agent performance across different task domains (web, OS, DB, embodied) based on the described domain split.
• Comparing planning strategies from different language models (GPT-5-mini vs. Claude-4-Sonnet) based on the history_gpt and history_claude fields.
• Analyzing task difficulty progression based on the described 'levels' (e.g., 5 levels for web tasks).
• Training or fine-tuning models for sequential decision-making based on the provided agent trajectory records.

Strengths

• Contains 786 total records across four distinct task domains.
• Includes trajectories from two different language model agents (GPT-5-mini and Claude-4-Sonnet).
• Web and embodied domains feature structured tasks with multiple difficulty levels (5 levels each).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality and structure require manual inspection after download.
• The provenance, collection method, and license for the underlying task data are unknown.

Provenance

Source

huggingface, author xwang2775

Collection Method

Likely contains agent trajectories and plans generated by language models (GPT-5-mini, Claude-4-Sonnet) executing tasks.

Freshness

Last updated 2026-04-17 02:22:21; freshness should be verified.