This repository contains the fine-tuning process of DistilGPT-2 for generating efficient n8n workflows based on pre-processed JSON datasets. The model is trained on a dataset of 331 input-output entries, enabling it to generate structured workflows based on text prompts.
The objective of this project is to fine-tune DistilGPT-2 on a dataset containing structured JSON workflows. The model learns to generate n8n-compatible JSON outputs from textual descriptions, helping automate workflow generation.
- Fine-tunes DistilGPT-2 using a structured dataset of n8n JSON workflows.
- Generates valid workflow JSONs from text inputs.
- Optimized for low-latency inference.
- Implements preprocessing, training, and validation pipelines.
- The dataset consists of 331 input-output pairs of text prompts and n8n workflow JSONs.
- Preprocessing includes tokenization and formatting for GPT-style training.
- Model: DistilGPT-2
- Training Data: Pre-processed n8n workflows
- Framework: Hugging Face Transformers
- Training Method: Supervised fine-tuning
- Optimizations: Gradient checkpointing, mixed precision training
- The fine-tuned model successfully generates structured n8n workflows.
- Accuracy and performance are evaluated using execution correctness.
- Enhance dataset with more complex workflows.
- Implement RLHF (Reinforcement Learning with Human Feedback) for better output quality.
- Integrate n8n API for real-time validation.
This project is licensed under the MIT License.
🔗 Contributions & Feedback
Feel free to open issues and contribute to the project!