WildClawBench-Trajectories

# WildClawBench Trajectories This dataset contains complete OpenClaw agent trajectories collected from the WildClawBench evaluation. The original packaged evaluation outputs remain available as `output_*.tar.gz`; `train.parquet` provides a table optimized for the Hugging Face Dataset Viewer, while `sessions/` provides individually browsable sessions for the Hugging Face Agent Trace Viewer. The current release contains 600 trajectories: 60 benchmark tasks evaluated with 10 models. ## Dataset Structure - `task_id`: WildClawBench task identifier. - `trajectory`: Full message sequence serialized once as a JSON array. - `model_name`: Evaluated model display name. - `task_category`: One of the six WildClawBench task categories. To keep Dataset Viewer rows small enough to load reliably, inline base64 image payloads are replaced by placeholders containing the original payload length and SHA-256 digest. Message order, image positions, MIME types, text, reasoning, tool calls, and tool results are preserved. The original image payloads and all task artifacts remain available in the corresponding source archive. ## Agent Trace Viewer The [`sessions/`](./sessions) directory contains one Pi session v3 JSONL file per model and task: ```text sessions/<model>/<task_id>.jsonl ``` Open any JSONL file and select the **Trace** tab to inspect the full session timeline, reasoning blocks, model responses, token usage, tool calls, tool arguments, and tool results. These trace files preserve the original inline image data; only the compact `trajectory` strings in `train.parquet` omit base64 image payloads. Each session header includes a `trace_status` field: - `completed`: the recorded execution ended cleanly. - `error`: the model returned an explicit error. - `interrupted`: one or more tool calls have no recorded result. For `error` and `interrupted` sessions, the Trace Viewer displays a final warning block. This block is explicitly labelled as a synthetic dataset-export marker; it does not replace or modify the original model and tool events. The current release contains 567 completed, 7 error, and 26 interrupted traces. ## Usage ```python from datasets import load_dataset dataset = load_dataset("internlm/WildClawBench-Trajectories") sample = dataset["train"][0] ```