Open Source Models & Infra
SynthCommonCorpus: The Largest Open Synthetic Pretraining Corpus Built for Agentic Models
Date, time, and room will be added once confirmed.
Talk overview
The SynthCommonCorpus is the largest open synthetic pretraining corpus for agentic models, designed to explore how models learn to act like they learn syntax. It includes tool calls, multi-step reasoning, and adversarial examples, addressing limitations in current data sources. The corpus is built using synthetic generation, ensuring traceability and enabling measurable memorization and contamination.
