shiqihe/WebVoyager-Trajectories-Qwen3.5-Omni

WebVoyager + GAIA Agent Trajectories (Qwen3.5-Omni)

Full browser-agent trajectories for 733 tasks (643 WebVoyager + 90 GAIA-web), produced by a browser-use agent driven by qwen3.5-omni-plus-2026-03-15 (multimodal, via Alibaba DashScope), run in a real headed browser. Every step records the exact LLM context (including the screenshot the model saw) and the action taken, plus a reference-grounded success verdict.

Results

Judged by the WebVoyager reference-grounded multimodal judge (qwen3.5-omni-plus-2026-03-15). Each task is the better of two independent runs (success-preferring merge — keep a run's trajectory if it succeeded, otherwise take the other):

split	success	rate
overall	511 / 733	69.7%
WebVoyager	457 / 643	71.1%
GAIA-web	54 / 90	60.0%

Layout

metadata.jsonl              one row per task (index; loaded by `load_dataset`)
run_summary.json            run + merge provenance and success counts
data/<source>__<id>/
  meta.json                 task id/site/source, question, start_url, reference answer, status, final answer
  tool_schema.json          structured-output (tool) schema used each step
  history.json              browser-use AgentHistoryList (structured actions/states)
  webvoyager_eval.json      success verdict + judge reasoning
  step_001/
    messages.json           EXACT LLM request context: [system, user(<agent_history> text, "Current screenshot:", image_url base64)]
    output.json             the LLM output for that step (thinking + action taken)
    screenshot.jpg          the page at the start of the step
    state.json              url + title at the start of the step
  step_002/ ...

messages.json is the verbatim chat-completion request (OpenAI-compatible) the agent sent that step — system prompt + a user message whose content is the agent history text, the literal Current screenshot:, and the screenshot as an inline base64 image (downscaled to 1280×720). It is preserved unchanged so a trajectory can be replayed/scored fully offline.

metadata.jsonl columns

id, source (webvoyager|gaia), site, question, start_url, reference_answer, reference_type, agent_answer, status (completed|timeout|error), num_steps, success (bool, judge verdict), judge_reasoning, used_reference, trajectory_dir

Usage

from datasets import load_dataset
meta = load_dataset("shiqihe/WebVoyager-Trajectories-Qwen3.5-Omni")["train"]   # the 733-task index
# full per-step trajectories are the files under data/<trajectory_dir>/

Notes

• Reference answers: WebVoyager's reference_answer.json (per site) and GAIA's inline Final answer; both are embedded per task and used to ground the success judge.
• Generated with browser-use; structured output via response_format json_schema; one screenshot per step (vision on).