qwen-coder-3b-n8n-sft

Qwen Coder 3B - n8n Workflow Generation (SFT)

A fine-tuned version of Qwen2.5-Coder-3B-Instruct specialized in generating n8n workflow JSON configurations from natural language descriptions.

Model Description

This model was fine-tuned using Supervised Fine-Tuning (SFT) with the TRL library to understand and generate n8n workflow structures. It can produce valid JSON workflow definitions including nodes, connections, and configuration parameters.

Intended Uses

Direct Use

• Generate n8n workflow JSON from natural language descriptions
• Assist developers in creating automation workflows
• Prototype workflow structures before manual refinement

Out-of-Scope Use

• Production-critical workflow generation without human review
• General-purpose code generation (use the base model instead)
• Non-n8n automation platforms

Quick Start

Installation

pip install transformers torch accelerate

Basic Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "eclaude/qwen-coder-3b-n8n-sft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Create an n8n workflow that fetches data from a REST API every hour and sends a Slack notification"

messages = [
    {"role": "system", "content": "You are an n8n workflow expert. Generate valid n8n workflow JSON configurations."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Streaming Generation

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "eclaude/qwen-coder-3b-n8n-sft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "Create a workflow to monitor a GitHub repository for new issues and create Trello cards"

messages = [
    {"role": "system", "content": "You are an n8n workflow expert. Generate valid n8n workflow JSON configurations."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

_ = model.generate(
    **inputs,
    max_new_tokens=2048,
    streamer=streamer,
    temperature=0.7,
    do_sample=True
)

Using with vLLM (Recommended for Production)

from vllm import LLM, SamplingParams

llm = LLM(model="eclaude/qwen-coder-3b-n8n-sft")

sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=2048
)

prompt = """<|im_start|>system
You are an n8n workflow expert. Generate valid n8n workflow JSON configurations.<|im_end|>
<|im_start|>user
Create a workflow to sync Airtable records with a PostgreSQL database daily<|im_end|>
<|im_start|>assistant
"""

outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text)

Quantized Inference (Low VRAM)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

model_id = "eclaude/qwen-coder-3b-n8n-sft"

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4"
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=quantization_config,
    device_map="auto"
)

# Use same inference code as above

Training Details

Training Data

The model was fine-tuned on eclaude/n8n-workflows-sft, a curated dataset of n8n workflow examples with natural language descriptions.

Training Procedure

• Framework: TRL SFTTrainer
• Infrastructure: Hugging Face Jobs
• Precision: Mixed precision (bf16)

Limitations

• JSON Validity: Generated workflows may require manual validation and adjustment
• Node Coverage: Performance varies by n8n node type; common integrations (HTTP, Slack, GitHub) perform better than niche nodes
• Complex Workflows: Multi-branch workflows with complex logic may be incomplete
• Version Drift: n8n workflow schema evolves; generated JSON may need updates for newer n8n versions
• No Execution Testing: Outputs are not tested against a live n8n instance

Recommendations

1. Always validate generated JSON before importing into n8n
2. Use as a starting point, not a complete solution
3. Review connection mappings and credentials placeholders
4. Test workflows in a staging environment first

Example Output

Prompt: "Create a simple webhook trigger that logs incoming data"

Generated Workflow:

{
  "name": "Webhook Logger",
  "nodes": [
    {
      "parameters": {
        "path": "webhook-log",
        "responseMode": "onReceived",
        "responseData": "allEntries"
      },
      "name": "Webhook",
      "type": "n8n-nodes-base.webhook",
      "typeVersion": 1,
      "position": [250, 300]
    },
    {
      "parameters": {
        "functionCode": "console.log(JSON.stringify($input.all(), null, 2));\nreturn $input.all();"
      },
      "name": "Log Data",
      "type": "n8n-nodes-base.function",
      "typeVersion": 1,
      "position": [450, 300]
    }
  ],
  "connections": {
    "Webhook": {
      "main": [[{"node": "Log Data", "type": "main", "index": 0}]]
    }
  }
}

Citation

@misc{qwen-coder-n8n-sft-2025,
  author = {eclaude},
  title = {Qwen Coder 3B - n8n Workflow Generation (SFT)},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/eclaude/qwen-coder-3b-n8n-sft}
}

Model Card Contact

For questions, issues, or contributions, please open a discussion on this repository or contact via Hugging Face.