Report
Project Greenhouse releases Gaggle model checkpoints in an open agentic search effort
The researchers say their reranker was trained from scratch on commonly available datasets, without a third-party backbone.
TLDR
In an October 8 paper, the Project Greenhouse researchers describe building a pointwise, decoder-only search reranker through pre-training from scratch and supervised fine-tuning, without using a third-party open-weight backbone. They say the bulk of their experiments used no more than a handful of GPUs. They released data, code, configurations and Gaggle model checkpoints, presenting the work as a first step toward fully open, sovereign agentic search.
Combined views
5
1 Source, first seen ago
17 reposts