2026/04/05
•地點: New York University (NYU)
Prompting and Fine-Tuning Qwen for Text-to-SVG Generation under Structured Output Constraints Authors/Creators
We study text-to-SVG generation as a structured sequence modeling problem in which a model must map natural language descriptions to valid, compact SVG markup. Our approach combines dataset analysis, SVG preprocessing, prompt design, and iterative model ablations under practical hardware constraints. We evaluate multiple configurations across AMD MI100 and Kaggle NVIDIA T4 environments, spanning Qwen2.5 instruction-tuned and coder-specialized model variants, different context lengths, LoRA settings, and training data scales. The strongest configuration uses Qwen2.5-Coder-3B-Instruct with a 3072-token context window and achieves 10.77991 on the private leaderboard and 13.11993 on the public leaderboard. Our results suggest that preprocessing quality, task-aligned prompting, and selecting a model family suited to code-like generation were more important than simply scaling baseline hyperparameters in isolation