Clawgym Trajectory: Long-Horizon AI Tool-Use Interactions

24,500 interaction trajectories were collected from ClawGym tasks using teacher rollouts from the MiniMax-M2.5 and GLM-5.1 models. The trajectories were filtered by verifier scores, resulting in a selection focused on long-horizon and tool-intensive scenarios. The dataset was authored by RUC-AIBOX and last updated on May 15, 2026.

Use Cases

• Training reinforcement learning agents for long-horizon planning based on the described multi-round trajectories.
• Benchmarking language model tool-use capabilities based on the average tool calls and tool types per trajectory.
• Analyzing the relationship between trajectory length and performance based on the average token count per interaction.
• Studying curriculum learning strategies based on the verifier-score filtering process mentioned in the description.

Strengths

• Contains 24,500 interaction trajectories, providing a substantial corpus for analysis.
• Trajectories are long-horizon, with an average of 13.00 rounds per interaction.
• Trajectories are tool-intensive, averaging 15.82 tool calls and 3.25 distinct tool types.

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

RUC-AIBOX

Collection Method

Collected from ClawGym tasks using teacher rollouts from MiniMax-M2.5 and GLM-5.1 models, then filtered by verifier scores.

Freshness

Last updated 2026-05-15 08:05:53; freshness should be verified.