WorkflowLLM is a data-centric framework designed to enhance LLMs' capabilities in workflow orchestration. The core of WorkflowLLM is WorkflowBench, a large-scale supervised fine-tuning dataset containing 106,763 samples across 1,503 APIs from 83 applications spanning 28 categories.
- [2024/10/29] The WorkflowBench dataset is live, empowering WorkflowLLM with the capability to orchestrate workflows across thousands of real-world APIs. Discover the structured training and evaluation scripts provided in this repository to enable end-to-end API orchestration.
The creation of WorkflowBench follows three main phases:
- Data Collection: We collect real-world Apple shortcuts from platforms like RoutineHub, transcribing them into Python-style code. Additionally, we enhance the workflows with hierarchical thought generation using ChatGPT.
- Query Expansion: ChatGPT is prompted to generate diverse and complex task queries to enrich the workflow dataset.
- Workflow Generation: An annotator model trained on the collected data generates workflows for synthesized queries, which are then quality-checked and merged with collected samples to form the final dataset.
Using WorkflowBench, we fine-tune Llama-3.1-8B to create WorkflowLlama, a model specifically optimized for workflow orchestration tasks. Experimental results demonstrate that WorkflowLlama excels at orchestrating complex workflows and generalizes well to unseen APIs.
In all, WorkflowLLM offers a powerful tool for automating sophisticated workflows, enhancing LLM-based solutions for real-world applications in process automation.
This dataset consists of two primary components: (1) transcribed real-world data, and (2) additional synthesized data. The distribution of action counts, workflow categories, and applications used across these two datasets is visually compared in the figure below:
You can download the complete dataset from the following link: Google Drive. Once downloaded, please unzip the contents into the directory ./data/.
We also included 50 samples in the ./data/data_samples.json file.
The folder structure under ./data/ is as follows:
./data/
│
├── dataset_split_keys.json
├── dataset_split_keys_ood.json
├── identifier2json.pkl
├── identifier2python.pkl
├── seed_data.json
├── statistics.pkl
└── synthesized_data.jsonHere are some descriptions for the data directory:
-
dataset_split_keys.json:
This file contains the dataset split for unseen instructions (In Distribution, ID). It defines how the data is divided based on new instructions that have not been seen during training. -
dataset_split_keys_ood.json:
Similar todataset_split_keys.json, but for unseen APIs (Out of Distribution, OOD). This file contains the split for instructions and APIs that are out of distribution, designed for testing how the model handles APIs that weren't seen during training. -
identifier2json.pkl:
A Python pickle file that stores API documentation in JSON format. The data is indexed by an API's identifier, and it is used to reference the APIs' descriptions, parameters, and other relevant details. -
identifier2python.pkl:
Another Python pickle file that stores API documentation but in Python-specific format. This data can be used to access the same API information, but formatted for Python usage (e.g., type hints, docstrings). -
seed_data.json:
This file contains the transcribed real-world data. This data serves as the "seed" data for building or augmenting the dataset. -
synthesized_data.json:
This file contains synthesized data generated to augment the dataset. The synthesized data helps increase the size and diversity of the dataset, ensuring that the model can generalize better. -
statistics.pkl:
A statistics file that contains summary information, such as the API categories used by each workflow, the number of actions, the number of nestings, and so on.
- Python 3.8 is recommended.
- All dependencies are listed in
./requirements.txtfor ease of installation.
pip install -r requirements.txtAdditionally, we used the DeepSpeed Zero Stage 2 training configuration. The specific configuration file can be found in ./configs/ds_config.json.
Given that the original Apple Shortcuts use a property list (plist) format, which is not human-readable and poses challenges for model training, we first convert the data into an Abstract Syntax Tree (AST) representation. This transformation enhances both the readability and the utility of the data for further processing. The pseudocode for the conversion algorithm is illustrated below:
By performing a pre-order traversal of the AST, we are able to obtain Python code corresponding to the Shortcuts' logic.
To convert .plist or .shortcut files into a Python-compatible format, execute the following script:
python ./preprocess/Convert_ShortCut_to_Python.py --folder_path {folder_path}After obtaining the transformed raw data, we used a prompt for ChatGPT to rename the variables and generate line-by-line comments for each workflow, along with the corresponding task plan and user query. The overall format of the data is as follows:
The format for each entry is shown in the figure below:
This repository contains tools for training and running inference with our model. Below, we provide instructions on how to begin training the model and running inference, along with useful configuration and troubleshooting information.
To start training, execute the following command:
sh ./scripts/train.sh {BASE_MODEL_PATH} {DATA_PATH}- {BASE_MODEL_PATH}: The path to the pre-trained model or an initial checkpoint to fine-tune. Ensure this path is valid and points to a compatible model checkpoint or architecture.
- {DATA_PATH}: The path to the dataset.
- The training process will automatically load the dataset, configure the model, and begin training.
- The script supports logging and saving intermediate checkpoints.
Once the model has been trained, you can run inference using the following command:
sh ./scripts/infer.sh {LOAD_PATH}- {LOAD_PATH}: The path to the trained model checkpoint. Ensure this points to a valid checkpoint directory or file that was saved during the training process.
| Model | BLEU (ID) | BLEU (OOD) | Weighted N-Gram (ID) | Weighted N-Gram (OOD) | AST (ID) | AST (OOD) | Data-Flow (ID) | Data-Flow (OOD) | Overall (ID) | Overall (OOD) | Pass Rate (ID) | Pass Rate (OOD) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Proprietary Models | ||||||||||||
| GPT-4o-mini | 0.4 | 0.4 | 1.5 | 1.6 | 29.5 | 29.5 | 37.0 | 36.3 | 26.8 | 26.5 | 54.8 | 47.5 |
| w/ ICL | 0.5 | 0.5 | 1.7 | 1.8 | 35.3 | 34.4 | 35.1 | 34.2 | 28.3 | 27.7 | 66.0 | 57.7 |
| GPT-4o | 0.5 | 0.4 | 1.8 | 1.7 | 33.5 | 31.8 | 37.3 | 36.9 | 28.5 | 27.7 | 56.6 | 47.5 |
| w/ ICL | 0.5 | 0.5 | 1.8 | 1.8 | 37.1 | 35.3 | 38.0 | 36.6 | 30.2 | 30.0 | 67.5 | 57.6 |
| Open-Source Models | ||||||||||||
| Qwen2-7B | 0.4 | 0.4 | 1.2 | 1.3 | 27.2 | 27.7 | 33.2 | 33.1 | 24.4 | 24.5 | 25.6 | 22.6 |
| w/ ICL | 0.5 | 0.5 | 1.2 | 1.3 | 30.2 | 29.8 | 32.4 | 32.9 | 25.2 | 25.3 | 28.2 | 26.4 |
| Llama-3.1-8B | 0.6 | 0.7 | 1.2 | 1.4 | 31.0 | 29.6 | 30.0 | 30.8 | 24.6 | 24.3 | 33.0 | 24.5 |
| w/ ICL | 0.7 | 0.7 | 1.3 | 1.4 | 34.0 | 32.4 | 32.6 | 32.4 | 25.3 | 25.2 | 40.2 | 32.7 |
| Llama-3.1-70B | 0.4 | 0.4 | 1.4 | 1.5 | 29.9 | 30.0 | 37.8 | 37.6 | 27.3 | 27.2 | 55.4 | 42.3 |
| w/ ICL | 0.4 | 0.4 | 1.6 | 1.5 | 34.1 | 32.9 | 39.1 | 38.4 | 29.5 | 28.7 | 67.6 | 61.4 |
| WorkflowLlama (8B) | 9.4 | 7.0 | 11.09 | 8.3 | 55.1 | 48.8 | 38.0 | 35.3 | 39.3 | 35.1 | 76.9 | 70.4 |
Feel free to us if you like WorkflowLLM.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.