BrowserSparseAttention — WebVoyager End-to-End Trajectories

Browser-agent trajectories for Qwen3-VL-8B running the WebVoyager benchmark under sparse-attention configurations. 60,567 files, ~2.4 GB (trajectories + per-step screenshots + observation JSONs).

Variants

Summarized set (8 variants × 200 tasks = 1,600 runs) — layout <variant>/webvoyager__<Site>--<id>/: dense, quest, block_sparse, tsa (default page size) and their *_b64 counterparts (page/block size 64).

Extra *_vortex runs — block_sparse_vortex (16,918 files) and quest_vortex (18,915 files) from the earlier vortex/flashinfer path, organized as <variant>_vortex/<timestamp>_b32|b64/webvoyager__<Site>--<id>/ (timestamped batch runs, batch size 32/64) with no run_summary.json. Present in the dataset but not summarized/graded.

Per-run layout

results/<variant>/webvoyager__<Site>--<id>/ contains:

• meta.json — question, golden reference_answer, start_url, final answer, status, num_steps
• history.json — full agent trajectory ({"history": [...]})
• step_NNN/ — screenshot (.jpg) + observation JSON per step
• tool_schema.json — browser tool schema

⚠️ Not a graded accuracy benchmark

status: completed only means the agent hit a Done state before the step/time cap (~90% of every variant, including dense, time out). No scorer compares answer to reference_answer. Use as a trajectory / screenshot corpus, not as an accuracy ranking. See the analysis folder in the code repo for details.