Pick the recipe
YOLO detection or segmentation, LLM LoRA / QLoRA SFT, image classification. Each shows only the fields it needs.
Fine-tuning
Fine-tuning Script Generator
Fill in the dataset path, model, epochs, batch size and learning rate and get a complete script: Ultralytics YOLO with data.yaml and train and val commands, a Hugging Face Transformers + PEFT LoRA script for chat SFT with TRL, or a torchvision/timm image-classification script. Each includes a requirements list, evaluation and export steps, and comments where a choice matters.
Loading the tool…
How it works
Fill in your numbers
Paths, model name, epochs, batch, learning rate, image size or sequence length, LoRA rank. Sensible defaults are pre-filled and explained.
Copy or download
train.py, the config file and requirements.txt as a zip, plus the shell commands to run them.
Formats, limits and the numbers that matter
YOLO defaults
yolo11n/s/m, imgsz 640, 100 epochs, batch 16, patience 20, then val and export to ONNX with the same imgsz.
LoRA defaults
rank 16, alpha 32, dropout 0.05 on q/k/v/o (and MLP projections optionally), lr 2e-4, cosine schedule with 3 percent warmup, 3 epochs, bf16, gradient checkpointing, packing off.
Classifier defaults
timm convnext_tiny or resnet50 pretrained, AdamW 3e-4, cosine, label smoothing 0.1, 20 epochs, ImageFolder layout from the labeler or splitter.
Common questions
Will it run as-is?
Yes on a machine with the listed requirements and the paths you gave; the script prints the resolved config at start so mismatches show immediately.
How do I choose the learning rate?
Start with the default, watch the loss for the first hundred steps: divergence means too high, no movement means too low. LoRA tolerates higher rates than full fine-tuning.
Can it target PaddleOCR?
A PaddleOCR detection and recognition config generator is next on the list.