Why this matters
Agent research increasingly depends on datasets that include not just prompts and responses but real tool calls and intermediate reasoning. This dataset supplies multi-turn trajectories where agents perform actual tool executions (terminal/browser/file ops), emit <think> chain-of-thought blocks, and record real tool responses — giving you grounded traces for training, fine-tuning, and diagnostic analysis.
What Sets It Apart
- Grounded tool execution: samples include real tool invocations and their actual responses (terminal commands, Playwright browser actions, file ops), not synthetic placeholders — so models can learn from realistic tool-result patterns and error conditions.
- Explicit reasoning traces: frequent and deep
<think>blocks (notably the Kimi config with large average<think>depth) expose intermediate chain-of-thought useful for SFT, rationale-conditioned training, and interpretability studies. - Two model-sourced configs: Kimi-K2.5 (7,646 samples) and GLM-5.1 (7,055 samples), together ≈14.7K samples with rich turn/tool-call statistics (hundreds of thousands of turns/tool-calls across the corpus), enabling cross-model comparisons and domain-specific subset selection.
Who it's for and trade-offs
Great fit if you want realistic training traces for tool-enabled agents — e.g., supervised fine-tuning (SFT) of tool-calling policies, developing agent skill libraries, or studying how chain-of-thought interacts with tool outcomes. The dataset’s Apache-2.0 license and standard parquet format make it easy to load with datasets/pandas/polars.
Look elsewhere if you need very large-scale web-scale conversational corpora (this is ~10K–100K in size) or multilingual coverage (this dataset is English-only), or if you require human-curated labels for downstream evaluation tasks rather than raw agent trajectories. Also note generation details reflect the source models and serving setups used during collection — treat model-specific artifacts accordingly when generalizing results.
