Awesome

Self-MAP

Setup

Use Python version <= 3.11, and install the required packages using the following command:

pip install -r requirements.txt

Usage

1. Download the MT-Mind2Web dataset and place it in the data/ directory.

2. Fill in the DATA_PATH and LOG_PATH variables as within the src/candidate_generation/conf/config.yaml and src/action_prediction/conf/config.yaml files.

   data:
       data_path: $(DATA_PATH)
       train_split_file: data/train/*.json
       test_split_files:
           test_task: data/test_task/*.json
           test_website: data/test_website/*.json
           test_subdomain: data/test_subdomain/*.json
   

   hydra:
       run:
           dir: $(LOG_PATH)
   

3. Set the environment variable OPENAI_API_KEY to your OpenAI API key.

Candidate Generation

Fine-tuning

python src/candidate_generation/train.py model=deberta-v3-base

Evaluation

Generate the ranks and scores for element candidates within the test-X set, where X can represent task, website, or subdomain.

python src/candidate_generation/evaluate.py\
    --model_path ${MODEL_PATH}\
    --data_path ${DATA_PATH}\
    --split_file data/test_${X}/*.json\
    --output_dir ${OUTPUT_PATH}\

Generate Pickle Files

Generate the pickle files for conversational action planning.

import pickle

def load_pickle(file_path):
    with open(file_path, 'rb') as file:
        data = pickle.load(file)
    return data

def write_pickle(data, file_path):
    with open(file_path, 'wb') as file:
        pickle.dump(data, file)

all_dict = {"scores": {}, "ranks": {}}
for file_path in [test_task_path, test_website_path, test_subdomain_path]:
    data = load_pickle(file_path)
    for key in ["scores", "ranks"]:
        for annotation_id, element in data[key].items():
            all_dict[key].setdefault(annotation_id, {}).update(element)

write_pickle(all_dict, output_path)

Conversational Action Planning

Fine-tuning

torchrun --nproc-per-node 4 --master_port=10086 \
    /src/action_prediction/train.py \
    model=flan-t5-base \
    train.per_device_train_batch_size=8 \
    train.gradient_accumulation_steps=1 \
    train.fsdp=False \
    train.num_gpus=4 \
    train.epoch=5 \
    run_id="full" \
    ++self_map.generation=False \
    ++self_map.memory_simplification=False \
    ++self_map.memory_refinement=False \
    ++self_map.multifaceted_matching=False \

Evaluation

python src/action_prediction/evaluate.py\
  +model_path=${MODEL_PATH}\
  model=flan-t5-base\
  +output_path=${OUTPUT_PATH}\
  +top_k=50\
  ++self_map.generation=False \
  ++self_map.memory_simplification=False \
  ++self_map.memory_refinement=False \
  ++self_map.multifaceted_matching=False \

Set self_map.${generation, memory_simplification, memory_refinement, multifaceted_matching} to True to enable the corresponding module.

LICENSE

Our code is derived from the Mind2Web project under the MIT License.
Our MT-Mind2Web dataset is made available under the CC-BY-4.0 license.

Citation

@inproceedings{self-map,
    author = {Deng, Yang and Zhang, Xuan and Zhang, Wenxuan and Yuan, Yifei and Ng, See-Kiong and Chua, Tat-Seng},
    booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
    doi = {10.18653/v1/2024.acl-long.477},
    editor = {Ku, Lun-Wei  and  Martins, Andre  and  Srikumar, Vivek},
    pages = {8795--8812},
    publisher = {Association for Computational Linguistics},
    title = {On the Multi-turn Instruction Following for Conversational Web Agents},
    url = {https://aclanthology.org/2024.acl-long.477},
    year = {2024}
}