Not even that hard: Greenhouse's from-scratch reranker trained on just a few GPUs
lintool · x · 2026-10-10
Jimmy Lin details Milestone 1 of Project Greenhouse: pre-training from ClimbMix then supervised fine-tuning on RLHN relevance data produced a competitive pointwise decoder-only reranker using just a few GPUs — "and it wasn't even that hard."
More from Research
- JevBench creator explains why open and API models get separate leaderboards: fairness — airesearch12 · 2026-10-11
- Harvard Med workshop teaches building AI co-scientists with ToolUniverse's 2,700+ tools — marinkazitnik · 2026-10-11
- RLVR easy wins are vanishing, scaling 'taste'-judged tasks is the next frontier — marktenenholtz · 2026-10-11
- ToolUniverse hits 2,900+ tools, lets AI agents run protein design campaigns on GPUs — marinkazitnik · 2026-10-11
- Mapping Für Elise to a Markov chain: music as audible mathematics — solyarisoftware · 2026-10-11
- Researcher finds many eval harness functions exist only to patch dataset flaws — remilouf · 2026-10-11