Explorer

This is the official codebase for Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents [ACL 2025 (Findings)]. This project is a collaboration between The Ohio State University and Microsoft Research.

• 🏠 Website
• 📖 Paper

Stay tuned for additional code releases and modules for this project.

📦 Training

Mind2Web-Live

cd train/
torchrun --nproc_per_node=4 train_qwen2vl.py --use_flash_attention --bf16 --train_dir <PATH_TO_SAVED_HUGGINGFACE_TRAJ_DATASET> --train_data_dir <PATH_ROOT_TO_RAW_TRAJS> --output_dir <OUTPUT_DIR> --num_train_epochs 10 --batch_size 64 --use-google-search

Multimodal-Mind2Web

cd train/
torchrun --nproc_per_node=4 train_qwen2vl.py --use_flash_attention --bf16 --train_dir <PATH_TO_SAVED_HUGGINGFACE_TRAJ_DATASET> --train_dir_order <EMPTY_DIR_TO_SAVE_ORDERED_HUGGINGFACE_TRAJ_DATASET> --train_data_dir <PATH_ROOT_TO_RAW_TRAJS> --output_dir <OUTPUT_DIR> --num_train_epochs 2 --batch_size 64 --model_name_or_path Qwen/Qwen2-VL-7B-Instruct --use-nogoto-gs-format --order_all_steps --learning_rate 1e-5

🧪 Evaluation

Mind2Web-Live

Step 1: Installation

conda create --name myenv python=3.12.5
pip install -r evals/mind2web_live_eval/requirements.txt

Step 2: Start x server and set the DISPLAY environment variable

Xvfb :99 -screen 0 1920x1280x16 &
export DISPLAY=:99
export OPENAI_API_KEY=xxxxxxxxxxxx

Step 3: Run the evaluation script:

python -m evals.mind2web_live_eval.evaluate_model    --index -1     --planning_text_model {qwen2-vl-7b|phi-3.5v}     --toml-path evals/mind2web_live_eval/configs/setting_qwen7b_40k_sample_10epoch_sync_1280_gs_filter.toml     --use-flash-attention --ckpt-path CKPT_PATH --temp 0.01 --log-dir LOG_DIR --viewport-width 1280

Multimodal-Mind2Web

To evaluate the performance of the trained model on the Multimodal-Mind2Web benchmark:

Step 1: Installation

conda create --name myenv python=3.12.5
pip install -r evals/mind2web_orig_eval/requirements.txt

Step 2: Download the DeBERTa candidate generation scores from the following link:

🔗 DeBERTa Score File

Step 3: Run the evaluation script:

cd evals
python -m mind2web_orig_eval.eval \
  --use-flash-attention \
  --ckpt-path <CKPT_PATH> \
  --log-dir <LOG_DIR> \
  --score-file <PATH_TO_DEBERTA_FILE> \
  --split {test_domain|test_task|test_website} \
  --model {qwen-7b|phi-3.5}

In-domain evaluation

Step 1: Installation

conda create --name myenv python=3.12.5
pip install -r evals/in_domain_eval/requirements.txt

Step 2: Set necessary env variables (OPENAI_API_KEY for evaluating API-based models)

export OPENAI_API_KEY=xxxxxxxxxxxx

Step 3: Run the evaluation script:

python -u -m evals.in_domain_eval.eval --input-file in_domain_test.json --ckpt-path <CKPT_PATH> --use-flash-attention --log-dir <LOG_DIR> --use-spiral

Structure of in_domain_test.json:

[
 <path to traj dir 1>,
 <path to traj dir 2>,
 ...
 <path to traj dir n>,
]

MiniWoB++

Step 1: Installation

conda create --name myenv python=3.12.5
pip install -r evals/miniwob/requirements.txt

Step 2: Run the evaluation script:

bash evals/miniwob/eval-explorer.sh

Citation

If you find this work useful, please consider starring our repo and citing our paper:

@inproceedings{pahuja-etal-2025-explorer,
    title = "Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents",
    author = "Pahuja, Vardaan  and
      Lu, Yadong  and
      Rosset, Corby  and
      Gou, Boyu  and
      Mitra, Arindam  and
      Whitehead, Spencer  and
      Su, Yu  and
      Awadallah, Ahmed Hassan",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.326/",
    pages = "6300--6323",
    ISBN = "979-8-89176-256-5",
}