Dev Mandal, a 20-year-old IIT Madras student, has released a dataset called 'computer-use-large' that's already surpassed 100,000 downloads on Hugging Face. The dataset captures human computer interactions — screenshots, mouse clicks, keyboard inputs — designed specifically for training AI agents to control computers like humans do.
This hits at exactly the right moment. Computer-use AI is the current gold rush, with Anthropic's Claude leading the charge and OpenAI rumored to be close behind. But these models are notoriously data-hungry, and quality datasets of actual human computer behavior are scarce. Most existing datasets are either synthetic, limited in scope, or locked behind corporate walls.
What's missing from the original coverage is crucial context about data quality and methodology. Without knowing how Mandal collected these interactions — were they crowdsourced? From real workflows? Privacy-sanitized? — it's impossible to judge whether this dataset will actually move the needle. The download numbers suggest developers are desperate enough for this type of data that they'll try anything, but downloads don't equal deployment success.
For developers building computer-use agents, this represents a rare opportunity to train on human behavior patterns rather than synthetic data. But proceed with caution — inspect the data quality thoroughly before committing training resources. The computer-use space is moving fast enough that a mediocre dataset could set your project back weeks.