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.
Related event: Sentence Transformers v6.0 Ships Late-Interaction Multi-Vector Models(27 posts)→
More from coding & agent
- Swarm's First Commercial Project: Autonomous Drones for Solar Parks in Greece — bittingthembits · 2026-08-18
- DeepSeek open sources agent runtime with multi-vendor orchestration — abhishek__AI · 2026-08-18
- Gemini Launches Managed Agents with Linux Sandboxes — _philschmid · 2026-08-18
- ZenML open-sources Kitaru: replay production agent traces against different models or prompts — htahir1 · 2026-08-18
- Opinion: Generated code is safer than package manager code — rickasaurus · 2026-08-18
- Went from writing 95% of my code to under 1% in less than a year — cross__entropy · 2026-08-18