Project Greenhouse builds an open agentic search reranker from scratch on a few GPUs
lintool · x · 2026-10-10
Jimmy Lin's group released Project Greenhouse (arXiv:2610.11922), showing a competitive pointwise decoder-only reranker can be pre-trained from scratch (ClimbMix) then fine-tuned (RLHN relevance data) with just a handful of GPUs — no third-party open-weight backbone. They release Gaggle model checkpoints plus data, code and configs for full independent reproduction, arguing fully open, sovereign agentic search models are feasible with modest compute.
More from coding & agent
- User pitches custom bot instructions, flags 500-char memory truncation — PawelHuryn · 2026-10-11
- AI-Generated Code Needs Sandboxes: From Linux Namespaces to microVMs — Abhishekcur · 2026-10-11
- CMU AI Agents course releases Deep Research lecture: building, evaluating and RL-training research agents — AkariAsai · 2026-10-11
- Every shows its Slack AI agent handling speaker contracts in a 4-step workflow — every · 2026-10-11
- HTMLSlides: free open-source AI skill turns prompts into presentable HTML decks — kamathsblog · 2026-10-11
- After automating 15+ real estate firms: the 5 agent workflows clients rave about — Warm-Reaction-456 · 2026-10-11