LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésMatériel1 min de lectureRédaction DisconfirmAI

Greenhouse Releases Open Reranker

Team releases Gaggle models trained from scratch, proving sovereign search components are feasible with modest compute.

2026-10-09 12:00 GMT+8

The Project Greenhouse team has released checkpoints, code, and data for their Gaggle series of reranking models, aiming for fully open and sovereign agentic search.

The approach avoids reliance on third-party pre-trained backbones, instead using a two-step recipe of pre-training from scratch followed by supervised fine-tuning. This was accomplished using only commonly available datasets and a handful of GPUs, challenging the norm where high-quality retrieval depends on closed or semi-open foundation models.

The authors report that their pointwise decoder-only reranker is competitive and have shared all artifacts to enable independent reproduction. These results are currently based on first-party reporting and have not yet been validated by external benchmarks.

Sources

  • arxiv.org↗
Continuer la recherche →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement