Public, high-quality agentic traces are a bottleneck for building and evaluating agents that use tools, interact with environments, and perform multi-step reasoning. This release raises the scale of openly available agent trajectories to ~1.7M complete trajectories, enabling broader training, imitation learning, and behavior analysis that were previously constrained by dataset size and heterogeneity.
What Sets It Apart
- Scale: ~1.7M rows assembled from 219 source datasets — roughly 4× the previous large open corpus — so you get far more diverse agent behaviors and failure modes in a single bundle.
- Heterogeneity of tasks and teachers: includes code repair, shell scripting, math problems, competitive programming and general computer-use tasks, with traces generated by many teacher models and labeled sources, which helps domain transfer and robustness testing.
- Terminus-2 / ShareGPT-style format: each row is a complete agent trajectory (messages list with roles user/assistant/tool), making the dataset directly usable by Harbor and other terminus-2 tooling for training, fine-tuning, or simulation.
- Preserved metadata: columns include original source, teacher model, optional reward signals and task identifiers — useful for filtering, evaluation splits, and reward-conditioned training.
Who It's For
Great fit if you want a large, open corpus of agent behaviors for imitation learning, offline RL, fine-tuning LLMs on tool-enabled trajectories, or studying failure modes across model families. The dataset is practical for researchers needing terminus-2-compatible examples and for engineers building evaluation suites that require many diverse agent trajectories.
Look elsewhere if you need human-authored conversation data (this is generated traces), privacy-filtered human logs, or datasets with consistent single-task schemas — source heterogeneity means some columns are null for subsets of rows and some traces reflect synthetic sandbox environments rather than real user interactions.
Where It Fits
Use this dataset to pretrain or fine-tune agents that must coordinate tool calls and reasoning steps, to build replay buffers for offline RL, or to construct large-scale benchmarks of agentic behavior. Combine it with smaller, human-verified corpora if you need human-ground-truth labels or clean user logs.
Data format and practical notes
- Format: parquet, terminus-2 conversation harness (messages as list[dict]).
- Metadata: original_source, original_teacher, optional reward, task_id and other preserved columns from sources.
- Origin: assembled with the Harbor framework; rows come from many synthetic and sandbox sources, so filter by original_source/original_teacher when creating evaluation or training splits to control for distributional bias.
