Datasets I've curated and published to support the ML community.
A focused dataset of Triton kernel bodies and CUDA→Triton translation pairs (a few thousand examples) for teaching models to write GPU kernels.
7.9k+ carefully filtered, high-aesthetic images collected for training and benchmarking diffusion or image-text alignment models.
Single-speaker Gujarati speech dataset featuring 8 hours of clean narration for ASR or TTS experiments.
Metadata-first dataset with sourced links to NSFW video content for research on content moderation or detection pipelines.
≈150 hours of multi-speaker Brazilian Portuguese recordings curated for high-quality text-to-speech synthesis.
25+ minutes of Barack Obama's voice samples from public speeches, optimized for RVC (Retrieval-based Voice Conversion) model training.