Your own models: fine-tune a model for your work

A small model fine-tuned for a specific task does that task almost as well as a large one, but with a fraction of the processing. Some examples:

• writing scripts in your style;
• describing photos of your products with your vocabulary;
• classifying your documents.

Genesis doesn't train models, because that takes a GPU and hours of compute. It does use any model in GGUF format that you train elsewhere.

#1. Prepare examples

Between 200 and 2000 question → ideal answer pairs are enough, in a JSONL file:

{"messages": [{"role": "user", "content": "30 s script for Origen specialty coffee"}, {"role": "assistant", "content": "SCENE 1 — …"}]}

Ideas for getting examples out of Genesis:

• Assistant conversations you liked;
• Library descriptions you corrected by hand;
• your own documents in the format you want.

#2. Train with Unsloth (free on Google Colab)

1. Open an Unsloth notebook for the base model you want.

2. Choose the base model for the task:

   Task	Base model
   Text	Qwen3 0.6B / 1.7B / 4B
   Vision	Qwen2.5-VL 3B or SmolVLM
   Search	EmbeddingGemma or Qwen3 Embedding

3. Load your JSONL and train with LoRA. With a few hundred examples it takes minutes on a T4 GPU.

4. Export to quantized GGUF; Q4_K_M is a good balance:

   PythonCopy

   model.save_pretrained_gguf("my-model", tokenizer, quantization_method="q4_k_m")

   For vision models also export the projector (mmproj-*.gguf). Unsloth's vision notebooks and llama.cpp's convert_hf_to_gguf.py --mmproj script generate it.

#3. Import it into Genesis

1. Go to Settings → Local AI → Import my model (GGUF).
2. Choose the file and the type (text, vision or search) and give it a name.

From the CLI:

genesis run local.import path=~/Models/my-model-Q4_K_M.gguf role=text name="My script model"

• The file stays where it is: Genesis just remembers it.
• From then on, the Assistant, documents, descriptions or search (depending on the type) use your model before the catalog ones.
• To stop using it, click the delete button on its card. Your file isn't deleted.

#Tips

• Start with the smallest model that does the task well; fine-tuning a 0.6B or 1.7B model is usually enough for narrow tasks.
• Set the effort slider to Minimum or Balanced: your fine-tuned model is still used, with less CPU and less memory.
• Check the result with the Ask a model tool (ai.chat with model=local/custom-…).