Anish13/web-agent-graph-dataset

Web Agent Graph Dataset

This dataset contains web navigation tasks in graph format with single-step negatives for training reward models.

Data Format

Each line in graph_dataset.jsonl is a JSON object with three main sections:

graph_data

Task metadata including domain, root_url, and goal description.

node_data

Current state information with URL, accessibility tree observation, and screenshot path.

step_data

Action and outcome with:

• action: The action taken
• obs: Next state observation
• step_number: Step index
• is_positive: True for gold actions, False for negatives
• negative_type: Classification (easy_negative, hard_negative, detour_negative)
• negative_reason: Explanation for negative actions

Usage

import json

# Load dataset
with open('graph_dataset.jsonl', 'r') as f:
    for line in f:
        entry = json.loads(line)
        goal = entry['graph_data']['goal']
        action = entry['step_data']['action']
        is_positive = entry['step_data']['is_positive']

Statistics

See conversion_stats.json for detailed statistics.

License

MIT License