HyWeBench

HyWeBench evaluates tasks that combine web search and browser operation. It contains 100 human-written tasks with fixed outcome checklists: 311 information items and 247 operation items. The checklist is shared across evaluated systems rather than regenerated for each trajectory.

Property	Released split
Split	test
Tasks	100
Starting mode	68 search / 32 operation
Switching pattern	68 A3 / 32 A4
Information checklist items	311
Operation checklist items	247
Language	English
License	Apache-2.0

Version 4

The 2026-11-11 release retains the 100 task IDs and the test schema. It revises 9 questions, updates time-sensitive checklist wording and evidence sites, and adds 2 information items, bringing the checklist to 558 items. Interpret relative dates in both the question and the checklist using the reference date recorded for that trajectory. Use fresh trajectories for the revised questions. Release manifest

Download

🤗 Dataset page · 📄 Direct JSONL download

from datasets import load_dataset

tasks = load_dataset("yzxjb/HyWeBench", split="test")
print(len(tasks))  # 100

For the existing JSONL runner:

hf download yzxjb/HyWeBench data/test.jsonl manifest.json evaluation/judge_v4.py \
  evaluation/relative_date_context.py --repo-type dataset --local-dir data/hywebench

Schema

Field	Meaning
task_id	Stable task identifier
question	Public task instruction
web	Initial browser URL; empty means start in search mode
domain	Task's associated site domain
pattern	Designed switching pattern (A3: search → operation; A4: operation → search)
start_mode	S for search or O for browser operation
checklist	Fixed items, each with id, goal, kind and site
dataset	HyWeBench

The policy receives the public question and initial URL. Keep the checklist as evaluation metadata; it is not part of the policy prompt. Preserve the empty web field on search-start tasks instead of constructing a URL from domain.

Run and evaluate

In a checkout of the 💻 HyWe code repository, configure the policy endpoint and judge as described in the 📘 evaluation guide:

source configs/hywe.env
TASKS=data/hywebench/data/test.jsonl
bash scripts/run_hywe.sh "$TASKS" runs/hywebench-v4 operation

HJ_OP_SEARCH_RULE=0 python data/hywebench/evaluation/judge_v4.py \
  evaluation/hywebench/hybrid_traj_judge.py \
  --run runs/hywebench-v4 --tasks "$TASKS" \
  --out runs/hywebench-v4-judge/verdicts.jsonl \
  --jpg-dir runs/hywebench-v4-judge/jpg --workers 6
python evaluation/hywebench/score.py runs/hywebench-v4-judge/verdicts.jsonl --tasks "$TASKS"

The v4 wrapper checks the task file and the upstream judge against the release manifest, then passes the recorded reference date to the judge. Use the downloaded v4 JSONL for both rollout and judging, and write v4 verdicts to a fresh output path.

The judge evaluates each checklist item against the recorded text and browser evidence. Task success requires every item; information and operation rates are micro-averages over the corresponding items. Missing judgments contribute zero achieved items in the full benchmark denominator.

Archived results

Model	Task success (%)	Information (%)	Operation (%)
Qwen3.6-35B-A3B base	28.0	74.8	43.7
HyWe-35B	47.0	82.2	68.8

These judgments use the earlier checklist revision. All 14 comparison and ablation verdict sets, the scoring script, and the protocol are available in the 📊 evaluation directory. Live pages and search results can change, so a new run may differ.

Related resources

🤗 HyWe-35B · 🤗 Hybrid SFT data · 🤗 Search SFT data