Clawgym Filtered: 15,777 Agentic AI Trajectories for Tool-Use Training

15,777 filtered and reformatted agentic trajectories derived from the RUC-AIBOX/Clawgym-Trajectory dataset. The data contains 342 million tokens according to the Qwen3-4B tokenizer and was generated using models like MiniMax-M2.5 and GLM-5.1 within the OpenClaw framework. The dataset was last updated on May 19, 2026, by author hyunwoongko.

Use Cases

• Training or evaluating language models on multi-step reasoning and tool-calling based on the described agentic trajectories.
• Studying the structure of chain-of-thought reasoning in agentic systems based on the reformatted 'reasoning_content' field.
• Benchmarking model performance on tasks requiring interaction with a suite of 19 tools as mentioned in the description.
• Analyzing data filtering and retention strategies for large-scale trajectory datasets based on the 64.4% retention rate from the original 24,497 rows.

Strengths

• Contains 15,777 curated trajectories, representing a 64.4% retention rate from the original raw data.
• Includes 342 million tokens of training data, as measured by a specific tokenizer (Qwen3-4B).
• Trajectories are structured with assistant chain-of-thought moved to a dedicated 'reasoning_content' field for clarity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale metrics like file size and specific column names are unknown.
• The dataset's source and generation method are described, but license information is unavailable, which may restrict use.

Provenance

Source

Filtered and reformatted from the RUC-AIBOX/Clawgym-Trajectory dataset on Hugging Face.

Collection Method

Generated by AI models (MiniMax-M2.5, GLM-5.1) using the OpenClaw framework with 19 tools.

Time Range

null

Freshness

Last updated 2026-05-19 08:50:08; freshness should be verified.

Geography

null