ACC: 10,802 Long-Context QA Pairs from Agent Trajectories

10,802 compiled long-context question-answer pairs derived from multi-turn agent trajectories, as introduced in the paper 'ACC: Compiling Agent Trajectories for Long-Context Training'. The dataset was created by groundhogLLM and was last updated on the platform in May 2026. It is sourced from trajectories of Search, Software Engineering (SWE), and SQL agents.

Use Cases

• Training language models for long-context reasoning based on compiled multi-turn trajectories.
• Fine-tuning models for tool selection and agent behavior based on evidence signals from agent trajectories.
• Benchmarking agent performance in domains like Search, Software Engineering, and SQL based on the described QA pairs.
• Studying methods for Agent Context Compilation (ACC) to improve supervision over scattered evidence in dialogues.

Strengths

• Contains 10,802 QA pairs, providing a substantial collection for training.
• Specifically designed for long-context training, addressing a noted gap in standard agent SFT.
• Derived from trajectories of three distinct agent types (Search, SWE, SQL), suggesting domain variety.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale details like file formats and sample data are unavailable.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

groundhogLLM, introduced in the associated research paper.

Collection Method

Compiled from multi-turn agent trajectories from Search, Software Engineering (SWE), and SQL agents.

Freshness

Last updated 2026-05-23 11:08:58; freshness should be verified.