FlowForge: Natural Language to n8n Workflow JSON

FlowForge is an open source tool that converts natural language descriptions into importable n8n workflow JSON using a fine tuned Gemma model trained entirely on AMD hardware. The problem is simple. Building n8n automations by hand takes time and requires knowing the platform's node structure by heart. General purpose language models are surprisingly bad at this task since n8n workflow JSON has a strict schema with node types connections and parameters that must all line up correctly. We built a full pipeline from scratch. Two thousand seven hundred eleven real n8n templates were collected cleaned to eighteen hundred and enriched three times using an LLM to produce natural language prompts for each workflow giving us fifty four hundred training examples split into forty five hundred ninety training rows and eight hundred ten held out rows for evaluation. Training ran on a single AMD Radeon PRO W7900 with forty eight gigabytes of VRAM using LoRA rank sixteen on Gemma 3 4B with plain PyTorch and PEFT since Unsloth segfaults on this RDNA3 card and Gemma 4 currently cannot be LoRA trained at all because PEFT does not support its custom linear layers. This alone took real debugging and is documented in the README as a known issue others can avoid. We measured results honestly using a three gate evaluation harness checking whether output is valid JSON whether it matches the n8n schema and whether it actually imports into a live n8n instance. The base Gemma 3 4B model produces zero valid workflows on held out prompts. Our fine tuned model produces valid schema compliant workflows twenty percent of the time and critically whenever it produces valid JSON that JSON is always schema correct meaning the remaining failure mode is simply cutting off on the longest workflows rather than generating the wrong structure. The full stack includes a FastAPI backend anyone can clone the repository run the training pipeline on their own AMD hardware and build on top of what we made.