CrowdGPT v2.6 adds security and collaboration features
Project author Vxtzq1 announced CrowdGPT v2.6 with server-side validation, cryptographically signed client releases, Byzantine cross-checks, improved installation and GPU selection, and a pipeline for collaborative dataset submissions. The project aims to collaboratively train an open 1-billion-parameter model using participants’ PCs. The author said v2.6 trained the model on 4.5 billion new tokens in six days. He also released CrowdBench, an open-source benchmark where people can add general-intelligence questions with deterministic answers.
The update directly addresses weaknesses in the earlier architecture, including fake or poisoned contributions, server overload and potentially malicious client code. The author said the checks make malicious contributions harder, while noting that they do not create a perfect federated-learning framework.