After disappointing zrank-4b and qwen-8b rerankers, team falls back to dsv4-flash
bo_wangbo · x · 2026-10-08
bowangbo reveals his team tried zrank-4b and a Qwen 8B reranker to filter false negatives while building the Q2D-Web dataset, found both disappointing, gave up on rerankers, and switched to dsv4-flash instead — a concrete data point behind his skepticism about the latest big rerankers.
More from Models
- Dev ditches OpenClaw for Grok bot, citing speed and free live X access — haltakov · 2026-10-08
- Developers fume as Google sits on Gemini 4 Argon, early access hinted — OfficialLoganK · 2026-10-08
- Glimmer model details: full training pipeline and 50 tok/s on a MacBook via 4-bit quant + DFlash — TimDarcet · 2026-10-08
- Unreleased OpenAI model produced nearly all research results from a single agent prompt — MikePFrank · 2026-10-08
- A Lee Sedol moment for math: AI proofs will be free for all, raging at OpenAI is pointless — RexDouglass · 2026-10-08
- OpenAI's published math proof closely matches unpublished draft a mathematician uploaded to ChatGPT — GeorgiaChal · 2026-10-08