DeepLearning_MidtermKaggle

Dataset Required Files

Place the following files in the working directory:

train.csv Columns: prompt, svg test.csv Columns: id, prompt

Data Preprocessing

The following filters are applied:

Validity Filtering SVG must start with Remove empty entries Structural Filtering SVG length ≤ 1600 ≤ 10 ≤ 8 ≤ 8 Token Filtering Token length ≤ 1700

Dataset Split Train / Validation split = 50% / 50%

Model Base model: unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit Parameters: ~0.9B Quantization: 4-bit LoRA Configuration Rank: 8 Alpha: 16 Target modules: q_proj, k_proj, v_proj, o_proj gate_proj, up_proj, down_proj

Inference Two-Stage Generation Strategy First Pass (Stable) temperature = 0.65 top_p = 0.90 max tokens ≈ 640 Second Pass (Fallback Attempt) temperature = 0.80 top_p = 0.95 max tokens ≈ 2000

If both fail → fallback SVG is used.

SVG Validation

Each SVG is checked using:

XML parsing (ElementTree) Root tag must be

Invalid outputs → replaced with fallback SVG

⏱ Runtime ~671.6 minutes for 1000 samples Sequential generation (no batching) Each sample may use up to 2 passes