Author: Ivan Aristy (NYU Tandon, CS-GY 9223 / ECE-GY 7123) Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct Best Public Score: 16.87 / 100 GitHub: ivanearisty/svg-gen

Models

Model	Type	Public Score	Description
componly-r32-adapter	LoRA adapter	16.87	Best model. Requires `merged-1.5b-r16` as base. LoRA r=32, alpha=64, trained on 45k competition-only samples.
merged-1.5b-r16	Full model	—	Qwen2.5-Coder-1.5B with Round 1 LoRA r=16 adapter permanently merged into weights. Base for the best adapter.
refined-7000	Full model	16.26	Full fine-tune from merged base, checkpoint 7000, CE loss 0.308. Standalone model.
r16-3epoch	LoRA adapter	15.47	Round 1 adapter. LoRA r=16, 3 epochs on 46k competition data. Load on `Qwen/Qwen2.5-Coder-1.5B-Instruct`.
mixed-r32-adapter	LoRA adapter	14.64	LoRA r=32, trained on 76k mixed data (competition + external). External data hurt performance.
codegen-1.5b	Full model	12.26	Code generation experiment — model outputs Python code instead of raw SVG.

Training Strategy

We use an iterative merge-and-retrain approach:

Round 1: LoRA r=16 on Qwen2.5-Coder-1.5B (46k competition data, 3 epochs)
Merge adapter into base weights
Round 2: Fresh LoRA r=32 on merged base (45k competition data, 2 epochs) — best model
Round 3: Merge Round 2 adapter, full fine-tune on DGX Spark

System prompt is critical: Removing it drops score from 53.8 to 18.8 (local ablation)
Less is more: Minimal system prompt ("Output valid SVG code only.") outperforms verbose prompts
LoRA > Full fine-tune: Despite lower CE loss, full fine-tuning scores worse than LoRA
Competition data only: External datasets (SVGX, OmniSVG) degraded performance
Greedy decoding optimal: Any sampling or penalty variation hurts structured SVG output