bfloat16 Sigmoid Saturates CrossEncoder Rankings; Fix Lifts nDCG From 0.18 to 0.68
tomaarsen · x · 2026-08-18
The fix the author most wants users to know about: CrossEncoder.predict now upcasts logits to float32 before the activation, because sigmoid in bfloat16 saturates and ties top candidates together, randomizing their order. NanoBEIR nDCG@10 jumps from 0.1849 to 0.6795. The release also tests 51 checkpoints directly — 29 text retrieval and 22 visual document retrieval, 17M to 8.8B parameters — covering LateOn, GTE-ModernColBERT, mxbai-edge-colbert, LFM2-ColBERT, colbertv2.0, answerai-colbert-small, ColPali, ColQwen and more.
More from coding & agent
- Agents turn conservative as context grows: edits shrink from 150 to 30 lines — dbreunig · 2026-10-16
- ai& becomes an official built-in Cline provider with 13 open-weight models — DavidBennett__ · 2026-10-16
- Open-source rizin-mcp lets LLMs run local binary and malware analysis via Rizin — Pure_Emergency1424 · 2026-10-16
- x402 as a free GTM channel: agents discover 100+ tools, no API keys needed — 0xJeff · 2026-10-16
- Sam Altman: stop getting better at prompts, build loops and graphs that prompt for you — goyalshaliniuk · 2026-10-16
- POV ski game on real Matterhorn elevation data, built with Claude Opus 5.5 — techartist_ · 2026-10-16