Synthetic Web Navigation Trajectories with Agent Actions

MolmoWeb-SyntheticTrajs contains synthetic web-navigation trajectories pairing instructions with webpage screenshot sequences and corresponding agent actions. The dataset, created by AllenAI, includes subsets like 'from_template' and 'task_seeded_wv' for training and evaluating web agents. It features text modality and is categorized as having between 100K and 1M data points.

Use Cases

• Train instruction-following agents using the provided instruction and agent action sequences for web navigation tasks.
• Evaluate the performance of multi-modal models on sequences of webpage screenshots paired with user instructions.
• Analyze patterns in synthetic agent actions like clicks and typing to understand simulated user behavior.
• Benchmark web automation agents across different dataset subsets such as 'from_template' and 'multi_agent'.

Strengths

• Contains multiple structured subsets including 'from_template', 'task_seeded_wv', and 'multi_agent' for varied experimentation.
• Data size is categorized between 100K and 1M entries, providing a substantial corpus for model training.
• Includes paired multi-modal data of text instructions and visual webpage screenshot sequences.

Limitations

• The dataset is entirely synthetic, which may not fully capture the complexity and noise of real-world human web navigation.
• Specific details on the number of rows, columns, and file formats are not provided, limiting precise technical assessment.
• Potential for bias exists in the synthetic generation process, which could affect model generalization to real user data.

Provenance

Source

AllenAI, hosted on Hugging Face.

Collection Method

Synthetically generated trajectories pairing instructions with webpage screenshots and simulated agent actions.

Freshness

Last updated on March 24, 2026.

Geography

Region is indicated as 'us', suggesting a United States focus or origin.