Yuanbao-Reranker: 0.6B/4B/8B Qwen3-based rerankers trained on synthetic preferences, open-sourced
_reachsumit · x · 2026-11-11
An arXiv paper presents Yuanbao-Reranker, a family of general-purpose dense rerankers built on the Qwen3 backbone at 0.6B, 4B and 8B scales, targeting search, recommendation and agentic RAG.
- Training data: answer-first, query-type-guided document-grounded synthetic queries with a three-stage curation funnel, yielding 2.75M 1:3 ranking groups
- Teacher score margins filter potential false negatives; negatives tiered into Hard/Medium/Moderate
- Joint loss of pointwise cross-entropy and a bounded minimum-gap hinge objective, plus cross-scale/same-scale distillation and layerwise adaptive merging
- Evaluated on CMTEB-R, MMTEB-R, BEIR, MLDR and MTEB-Code
Weights are open-sourced on Hugging Face.
More from Models
- MiniMax H3 open ecosystem mapped: 49 family groups, 63 descendant nodes since August — cocktailpeanut · 2026-11-11
- Microsoft patent reveals real-time AI speech rewriting with lip-synced live video — Significant_Mail5398 · 2026-11-11
- Claude Code's web search tool has a built-in classifier that refuses some searches — NathanpmYoung · 2026-11-11
- Indie dev gnukeith teases free self-trained LLM landing mid-January with a companion alignment model — gnukeith · 2026-11-11
- Indie trainer gnukeith bets on sub-9B models: a 2B model can likely do 75% of your tasks — gnukeith · 2026-11-11
- Leaked chart claims Anthropic's internal Model 4 hit 178.76 AECI — haider1 · 2026-11-11