Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge
Paper • 2506.21506 • Published • 52
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Training selections of raw web-agent trajectories built on public benchmark tasks (see per-source licenses). Access is reviewed manually. Please describe your intended use.
Log in or Sign Up to review the conditions and access this dataset content.
Web Trajectories SFT
Supervised fine-tuning selections built from lukakuma/web-trajectories (judged-successful, completed sessions of a DOM-based, tool-calling web agent). One row = one LLM call: the exact input (messages + tools) and the exact output (response). The agent's own calls of every kind are included (actions, planning tools, goal decomposition, completion check), because the target model is meant to take all of them over.
Configs
| config | what is in train |
tasks | sessions | rows (LLM calls) | rows with train_target=false |
|---|---|---|---|---|---|
cap1 |
best session per task | 496 | 496 | 14,986 | 379 |
cap3 (default) |
best + up to 2 more that took a different path | 496 | 1,291 | 45,680 | 3,145 |
all |
every session that passed the filter | 496 | 7,187 | 265,946 | 22,664 |
test split (shared) |
held-out tasks, picked like cap3 |
85 | 234 | 8,250 | 649 |
Four training variants come out of three configs:
| variant | config | loss |
|---|---|---|
| A | cap1 |
all assistant outputs |
| B | cap3 |
all assistant outputs |
| C | all |
all assistant outputs |
| D | cap3 |
skip outputs of rows with train_target = false (the turn's tool call failed, or it repeated the previous call exactly) — the row stays in context |
Index configs: tasks (every task with split and held_out_by) and sessions (each selected session with its pick rank and the quality numbers used to rank it).
from datasets import load_dataset
import json
train = load_dataset("lukakuma/web-trajectories-sft", "cap3", split="train")
test = load_dataset("lukakuma/web-trajectories-sft", "cap3", split="test")
row = train[0]
messages, tools, target = json.loads(row["messages"]), json.loads(row["tools"] or "[]"), json.loads(row["response"])
use_loss = row["train_target"] # variant D only
Columns
| column | type | meaning |
|---|---|---|
session_id, turn_index |
str, int | session (join key to the raw repo) and call order |
segment_id |
int | consecutive rows of a session whose inputs extend each other (see Packing) |
source, task_id, domain, category, language |
str | benchmark, task, start site, original stratum, instruction language |
purpose |
str | act, decompose_goal, plan_tool:*, completion_verify, text_only …; null where the source session had no purpose log |
messages |
json str | exact model input (system prompt, task, history, tool results, one-off hint) |
tools |
json str | tool definitions offered on this call; null for calls without tools |
response |
json str | exact model output blocks (text, tool_call with name and JSON input) |
train_target |
bool | false if this turn's tool call failed or exactly repeated the previous call |
action_ok |
bool | every tool call of this turn returned success (null = unknown) |
in_tokens, out_tokens |
int | usage recorded at run time (agent tokenizer) |
role |
str | act / plan / report / decompose / verify / text, read from the call's content (see below) |
tool_names |
list[str] | tools called in response, in order |
Call roles
role is derived from each call's content, so it is filled for every row (the runner's own purpose label is missing for about a quarter of rows and files report writing under act or text_only depending on the version). Where purpose exists the two agree on 95.7% of rows; the rest is almost entirely report writing.
| role | rule | share |
|---|---|---|
act |
an environment tool was called (planning calls may be batched with it) | 77.7% |
plan |
only plan tools: create_plan, update_subtask_state, finish_subtask, finish_plan, view_subtasks |
12.9% |
report |
write_report (optionally with plan tools) |
4.1% |
decompose |
no tools offered; system prompt starts “You are a planning assistant” | 2.7% |
verify |
no tools offered; system prompt is the completion auditor | 2.6% |
text |
tools offered, none called | 0.1% |
To train on one role only, filter rows by role (e.g. decompose alone, or everything except verify).
How the selections were made
- Test tasks first. Whole start sites with at most 5 tasks were held out until about 5% of tasks (English web sources only; 32 tasks), then every 10th task per source in category order (57 tasks). Test tasks never appear in any
trainsplit. - Session filter (applied to train and test): drop a success that is the only one on a task whose success rate is below 20% (likely judge error), sessions with more than 2x the task's median action turns, and sessions with more than 25% failed actions.
- Per task. Rank by failed-action share + repeated-call share, then closeness to the task's median length.
cap1takes the first;cap3adds up to two more by farthest-point selection on the set of tools and URLs used (Jaccard distance ≥ 0.2). - Site balance (
cap1,cap3train): at most 30 sessions per start site; later picks are dropped first, never a task's first pick. allkeeps every filtered session (no cap, no balance).
Packing
Within a segment_id, each row's input is the previous row's input minus its trailing one-off <system-hint>, plus the previous response and its tool results. Rows can be trained one by one (exact), or a segment can be packed into one sequence (the last row's messages + response, loss on every assistant turn); packing drops the one-off hints the earlier rows saw unless they are kept with attention masking. cap3 has 9,143 segments for 45,680 rows. Inputs run up to about 107k tokens (agent tokenizer), so a 128k context holds every row.
Caveats
- Everything listed in the raw repo's card applies (single judge, scrub rules, third-party page text).
- Test tasks come from the same public benchmarks; they are held out from training here, not unseen by the agent that produced them.
- Prompt versions: the action-turn policy is one version for all English sessions and one for the Korean
canvassessions; the tool list varies by 1–4 tools over the period (54 tools are in every version). canvas(Korean) is small; it is not up-weighted.
Licenses and attribution
| source | origin | license | train rows (cap3) |
|---|---|---|---|
om2w |
Online-Mind2Web | CC-BY-4.0 | 28,484 |
webvoyager |
WebVoyager | Apache-2.0 | 7,870 |
webbench |
WebBench | MIT | 7,451 |
canvas |
In-house canvas tasks (Korean instructions) | CC-BY-4.0 | 1,279 |
mind2web2 |
Mind2Web 2 (dev set) | CC-BY-4.0 | 596 |
Please cite the original benchmarks: Online-Mind2Web (Xue et al., 2025, arXiv:2504.01382), WebVoyager (He et al., 2024), WebBench (Halluminate), Mind2Web 2 (Gou et al., 2025, arXiv:2506.21506).
- Downloads last month
- -