TinyVLM-470M
MultimodalA 470M parameter Vision-Language Model built from scratch with LLaMA 4-style iRoPE, Grouped Query Attention, and Mixture of Experts. Trained on a single RTX 3090 in ~35-40 hours, achieving 64.5% preference accuracy.
Things I've built from scratch
A 470M parameter Vision-Language Model built from scratch with LLaMA 4-style iRoPE, Grouped Query Attention, and Mixture of Experts. Trained on a single RTX 3090 in ~35-40 hours, achieving 64.5% preference accuracy.
Controlled benchmark of classical vs diffusion augmentation for few-shot satellite image classification. Fine-tuned SD 1.5 LoRA on EuroSAT, showing selective synthetic augmentation closes to within 0.4% of real-only performance at 64-shot.
GPT-style decoder-only transformer implemented from scratch in PyTorch with custom BPE tokenizer, KV-cache optimization, and a training loop with mixed precision, gradient clipping, and early stopping.
Chrome extension that hides spoiler metadata (difficulty, hints, topics, company tags) on LeetCode and NeetCode problem pages for distraction-free practice.