Fine-tuning LLM for n8n

I’m sure I’m not the first thinking about it, is there a project to fine tune a LLM on all the n8n templates available ? So we we could build workflows only with prompt.

You're *definitely* not the first—this is an idea that’s been circulating in AI + automation communities for a bit, and for good reason. Here's a breakdown of what's been happening and how you could jump in: --- ### 🔍 Has it been done? There isn't a **widely adopted open-source project** *yet* that fine-tunes a local LLM specifically on **all n8n workflows/templates** to enable full prompt-based automation. But several adjacent efforts are worth noting: #### 1. **LangChain + n8n integrations** Some devs have created wrappers where you use LangChain-style prompting to dynamically build n8n workflows (or parts of them). But these usually rely on **inference-based prompting**, not actual **fine-tuned LLMs**. #### 2. **Prompt Engineering Workflows** Some folks have created GPT-based agents that understand the **structure** of n8n JSON templates and can generate full workflows on the fly — but again, this is zero-shot or few-shot prompting. Not fine-tuning. #### 3. **n8n community experiments** There are a couple of posts in the [n8n community forum](https://community.n8n.io/) where people discuss having ChatGPT or local LLMs generate workflows — but these are generally scripted, not model-trained. --- ### 🧠 What would a fine-tuning project look like? A proper LLM finetune on n8n templates would need: - **Data**: A dataset of thousands of `.json` n8n templates (workflows) - **Input–Output Pairs**: Natural language prompts paired with the corresponding n8n JSON - **Tokenizer support**: Possibly specialized tokens for nodes, parameters, connections - **Finetuning infra**: LoRA or QLoRA on something like **Qwen 1.5**, **Mistral**, or **TinyLlama** if you’re going local - **Agent layer**: Optionally a wrapper that does auto-debugging or validation with the n8n API --- ### 🤖 Why hasn’t it taken off yet? - **Template diversity**: n8n workflows are extremely variable depending on nodes, credentials, APIs, custom JS - **Security**: Some include sensitive keys or environment variables - **Lack of standard prompt-schema pairs**: There's no massive public dataset pairing prompts with outputs like “Build me a Slack-to-Notion sync” → workflow.json --- ### 🧰 You could build this! (Seriously.) If you're down to tinker: 1. **Scrape templates**: Grab all public n8n templates from: - [n8n.io/templates](https://n8n.io/workflows) - GitHub search (`language:json n8n`) - Community forums 2. **Build prompt-data pairs**: - Use ChatGPT or a local model to *generate prompts* that describe each template - Format as: `{ "prompt": "Create a webhook that sends email via SendGrid", "workflow_json": {...} }` 3. **Train**: - Use LoRA/QLoRA finetuning with tools like [Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) or [SFTTrainer from HuggingFace](https://huggingface.co/docs/transformers/main/en/sft) 4. **Wrap in agent**: - Build a small interface (like a CLI or WebUI) where users prompt, get JSON, and push to n8n via their REST API. --- ### 🧩 Bonus idea Use **Structured Output prompting** (like JSON schemas) or [function calling-style APIs](https://platform.openai.com/docs/guides/function-calling) to generate only parts of a workflow and incrementally build with user confirmation. --- Would you want help starting a repo or getting a prompt-data generator going? You could make a dope niche project out of this that’s useful *and* publishable.

now rewrite in my style not using lists or bullet points but in paragraph style with no bold or italic or dashes or hyphens

You're definitely not the first to think about it, and it's a solid instinct. A lot of people in the AI and automation scenes have floated this idea—letting someone describe a workflow in plain language and having a model generate the n8n workflow JSON in response. Right now, though, there isn't a widely adopted open-source project that has actually fine-tuned a local LLM on the full set of n8n templates. Some experiments exist, mostly relying on prompting large models like GPT to generate workflow JSONs based on text descriptions. These are more like inference-based setups, not true fine-tuned models. There are a few LangChain wrappers and custom agents floating around that can dynamically build workflows, but again they depend on clever prompting rather than training a model directly on the templates. If you poke around the n8n community forums, you’ll find a handful of posts where folks talk about using ChatGPT or other models to generate n8n workflows, but none of them have taken it to the level of fine-tuning. There’s also not really a public dataset out there that pairs a human-readable prompt with the actual n8n workflow JSON. That’s probably the biggest reason this hasn’t taken off—n8n templates can be pretty complex and vary wildly depending on what services they connect to, what credentials they require, or how much custom code is involved. Still, it’s doable. If someone wanted to take it on, the workflow is pretty straightforward in theory. You’d gather a bunch of templates from the official n8n site or scrape GitHub, then come up with prompts that describe what each template does in natural language. Those prompts could be generated manually or with a model helping out. Once you’ve got enough pairs, you could fine-tune something like Mistral or Qwen using a tool like Axolotl or HuggingFace’s SFTTrainer. From there, it’s just a matter of building a small interface—maybe a simple WebUI or CLI—that takes in a prompt, runs it through the model, and returns a workflow JSON that can be imported into n8n. Optionally, you could also have it hit the n8n API directly. A cool twist on this would be using structured output prompting, like giving the model a partial schema for the workflow and asking it to fill it in step by step. That kind of iterative generation might actually be more reliable than having the model guess the entire thing at once. You could even build a validating agent that checks each generated node before adding it to the graph. If you're interested, I’d love to help you start building this out—there’s a lot of potential here for a very real and useful open-source project.

write a quick outline of how you would build it in the same style and method as you composed the previous response

To build this, I’d start by collecting as many public n8n templates as possible. The official n8n site has a big list of workflows and GitHub is another great place to mine for user-created templates. Once I had a solid set of those, I’d write or generate natural language prompts that describe each one. This could be something like “create a webhook that triggers when a new row is added to Google Sheets and sends a Slack message,” paired with the actual JSON workflow. These pairs would form the core of the training data. Next, I’d pick a model that’s lightweight enough to run locally but powerful enough to handle structured output. Something like Mistral or Qwen would be a good fit. I’d format the data into a simple instruction format with input as the prompt and output as the workflow JSON, and then fine-tune the model using something like Axolotl or HuggingFace’s supervised trainer. Once the model is trained, I’d set up a minimal interface—probably a small Python or Node.js backend with a frontend where someone can type a request and see the generated workflow. To make it more interactive, I’d add validation so the system checks if each part of the workflow is valid JSON and uses real n8n node types. Eventually I’d want to plug it directly into the n8n API so it could create and update workflows on the fly. Over time, I’d refine the dataset, improve prompt clarity, and maybe even create a feedback loop where users can rate or correct the outputs to improve the model. The whole goal would be to make building automations feel like chatting with a very competent builder who understands both natural language and the structure of n8n.