Benchmark Dataloader
A small benchmarking harness for multimodal dataloaders, built to surface throughput bottlenecks before they become expensive training-time surprises.
San Francisco, CAMember of Technical Staff at Datology
Building multimodal systems where benchmarks can change the next decision.
At Datology, I work on evaluation, curation, and training systems for multimodal models. I want benchmark results, pipeline choices, and training outcomes to agree.
Field note
Good research infrastructure shortens the distance between a data choice and a trustworthy comparison.
Research tile
DatBench
A benchmark designed to make VLM evaluation more discriminative and decision-useful.
Read paper
Personal anchor
Haoli Yin
Research engineer focused on multimodal curation, evaluation, and training systems.
Current work
What I'm working on now
Three lanes: evaluation, curation, and training systems. The point is to make the next experiment clearer and harder to fool.
Decision-useful VLM evaluation
Designing benchmarks and scoring paths that are selective enough to inform curation and model choices, not just report a leaderboard.
Multimodal curation and data quality
Working on embeddings, filtering, and dataset export paths for large image-text corpora, with attention to alignment and redundancy.
Training and launch systems
Building the dataloading, evaluation, and multi-node launch infrastructure that makes multimodal iteration fast enough to matter.
Public artifact
A smaller public tool, but the same bias toward practical research infrastructure.