internlm/WildClawBench-Trajectories

WildClawBench Trajectories

Complete OpenClaw agent trajectories from the WildClawBench evaluation — every message, reasoning block, tool call, and tool result from real long-horizon agent runs, released for independent verification, side-by-side comparison, and trace-level analysis.

Each evaluated model covers the full 60-task suite, and the collection is continuously updated as new models join the leaderboard. The directories under sessions/ always reflect the current model roster.

The same data is provided in three forms:

Form	Optimized for	Images
train.parquet	Hugging Face Dataset Viewer, load_dataset	Replaced by hash placeholders
sessions/<model>/<task_id>.jsonl	Hugging Face Agent Trace Viewer	Original inline data preserved
output_<model>.tar.gz	Raw evaluation outputs (scores, usage, logs, agent-produced files)	Original files preserved

The WildClawBench Family

Repository	What's inside
WildClawBench	The benchmark itself: task data and Docker images for all four harnesses, run via the official pipeline
WildClawBench-Harbor	All 60 tasks in Harbor format — evaluate any Harbor-supported agent with a single harbor run
WildClawBench-Trajectories (this repo)	Complete trajectories from our frontier-model evaluations

Dataset Structure

Each row of train.parquet is one (task, model) run:

• 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 output_<model>.tar.gz archive and in the sessions/ trace files.

Agent Trace Viewer

The sessions/ directory contains one Pi session v3 JSONL file per model and task:

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.

Raw Evaluation Outputs

Each output_<model>.tar.gz archive contains the per-task outputs exactly as generated by the evaluation pipeline: per-metric scores (score.json), token usage and cost (usage.json), execution logs, the full conversation trace, and all files the agent produced (task_output/).

hf download internlm/WildClawBench-Trajectories output_claude_fable5.tar.gz --repo-type dataset --local-dir .
tar -xzf output_claude_fable5.tar.gz

Usage

from datasets import load_dataset

dataset = load_dataset("internlm/WildClawBench-Trajectories")

# One (task, model) run per row
sample = dataset["train"][0]
print(sample["task_id"], sample["model_name"], sample["task_category"])

# All trajectories for a given model
runs = dataset["train"].filter(lambda x: x["model_name"] == "Claude Fable 5")

# Parse the message sequence
import json
messages = json.loads(sample["trajectory"])

Citation

If you use these trajectories in your research, please cite WildClawBench:

@article{ding2026wildclawbench,
  title={WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation},
  author={Ding, Shuangrui and Dai, Xuanlang and Xing, Long and Ding, Shengyuan and Liu, Ziyu and JingYi, Yang and Yang, Penghui and Zhang, Zhixiong and Wei, Xilin and Fang, Xinyu and others},
  journal={arXiv preprint arXiv:2605.10912},
  year={2026}
}