Training Summary
- Base model:
unsloth/Qwen3.5-4B
- Dataset:
yoavf/svg-animal-illustrations
- Training split:
train
- Examples: 1,416
- Epochs: 1
- Max sequence length: 2048
- Batch size: 1
- Gradient accumulation: 8
- LoRA rank: 32
- LoRA alpha: 32
- Trainable parameters: 42,467,328
- Runtime observed: about 25.2 minutes on NVIDIA RTX PRO 6000 Blackwell Workstation Edition
- Final train loss: 0.01996
Evaluation Setup
A qualitative comparison was generated on 20 prompts: 10 animal SVG prompts and 10 non-animal SVG prompts.
Outputs were generated from both the base model and the trained LoRA adapter. The comparison artifacts are included in comparison-20/.
comparison-20/comparison.json: raw prompt/output/metric datacomparison-20/comparison.html: standalone visual comparison report
Aggregate SVG Checks
| Model | outputs with `` when generation is capped. The visual HTML report uses sandboxed previews so malformed SVG cannot break the viewer.
Per-Prompt Results
| # | Category | Prompt | Base `` | Base chars | Trained chars |
|---:|---|---|---:|---:|---:|---:|---:|---:|
| 1 | animal | Create a cute blue rabbit SVG illustration. Output SVG only. | yes | no | yes | no | 972 | 876 |
| 2 | animal | Generate a playful orange cat as SVG code only. | yes | no | yes | no | 820 | 828 |
| 3 | animal | Draw a green parrot vector illustration in SVG format. | yes | no | yes | no | 1207 | 947 |
| 4 | animal | Make a small brown dog with big eyes as SVG only. | yes | no | yes | no | 904 | 771 |
| 5 | animal | Create a pink pig sitting happily as SVG code only. | yes | no | yes | no | 883 | 890 |
| 6 | animal | Generate a yellow duck SVG illustration. Return only the SVG. | yes | no | yes | no | 938 | 846 |
| 7 | animal | Draw a gray hamster with round ears in SVG format. | yes | no | yes | no | 966 | 912 |
| 8 | animal | Make a black-and-white cow cartoon as SVG code only. | yes | no | yes | no | 862 | 867 |
| 9 | animal | Create a red fox curled up as an SVG illustration. | yes | no | yes | no | 1134 | 788 |
| 10 | animal | Generate a purple fish with fins as SVG only. | yes | no | yes | no | 1254 | 796 |
| 11 | non_animal | Create a red sports car SVG illustration. Output SVG only. | yes | no | yes | no | 915 | 806 |
| 12 | non_animal | Generate a blue coffee mug as SVG code only. | yes | no | yes | no | 789 | 825 |
| 13 | non_animal | Draw a simple green house with roof in SVG format. | yes | yes | yes | no | 1052 | 854 |
| 14 | non_animal | Make a yellow sun over mountains as SVG only. | yes | no | yes | no | 1064 | 869 |
| 15 | non_animal | Create a purple rocket ship SVG illustration. | yes | no | yes | no | 954 | 814 |
| 16 | non_animal | Generate a desk lamp icon as SVG code only. | yes | yes | yes | yes | 658 | 816 |
| 17 | non_animal | Draw a colorful birthday cake as SVG format. | yes | no | yes | no | 846 | 993 |
| 18 | non_animal | Make a city skyline at night as SVG only. | yes | no | yes | no | 1201 | 907 |
| 19 | non_animal | Create a laptop computer SVG illustration. | yes | no | yes | no | 1173 | 930 |
| 20 | non_animal | Generate a beach umbrella and waves as SVG code only. | yes | no | yes | no | 891 | 940 |
Animal vs Non-Animal Coverage
- Animal prompts: 10
- Non-animal prompts: 10
Visual Comparison
The table below uses rendered PNG previews. If a generated SVG could not be rendered, the PNG shows a Render failed placeholder instead of a broken image icon.
Table with columns: #, Category, Prompt, Base model, Trained LoRA| # | Category | Prompt | Base model | Trained LoRA |
|---|
| 1 | animal | Create a cute blue rabbit SVG illustration. Output SVG only. | | |
| 2 | animal | Generate a playful orange cat as SVG code only. | | |
| 3 | animal | Draw a green parrot vector illustration in SVG format. | | |
SVG Checks
| # | Base has `` |
|---:|---:|---:|---:|---:|
| 1 | yes | no | yes | no |
| 2 | yes | no | yes | no |
| 3 | yes | no | yes | no |
| 4 | yes | no | yes | no |
| 5 | yes | no | yes | no |
| 6 | yes | no | yes | no |
| 7 | yes | no | yes | no |
| 8 | yes | no | yes | no |
| 9 | yes | no | yes | no |
| 10 | yes | no | yes | no |
| 11 | yes | no | yes | no |
| 12 | yes | no | yes | no |
| 13 | yes | yes | yes | no |
| 14 | yes | no | yes | no |
| 15 | yes | no | yes | no |
| 16 | yes | yes | yes | yes |
| 17 | yes | no | yes | no |
| 18 | yes | no | yes | no |
| 19 | yes | no | yes | no |
| 20 | yes | no | yes | no |
Usage
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = 'unsloth/Qwen3.5-4B'
adapter = 'YOUR_USERNAME/qwen35-svg-animal-lora'
tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base_model, trust_remote_code=True, device_map='auto')
model = PeftModel.from_pretrained(model, adapter)
Limitations
- This is a LoRA adapter, not a merged full model.
- The dataset is small, so the adapter may overfit to simple animal illustration patterns.
- Generation quality should be judged visually with the included comparison report.